Editors
Jan-Ola Östman Kyoko Ohara
University of Helsinki, Keio University, Japan
Finland
Advisory Board
Peter Auer Mirjam Fried Knud Lambrecht
University of Freiburg, Germany Charles University, Prague, University of Texas at Austin, USA
Czech Republic
Hans C. Boas Michael Tomasello
University of Texas at Austin, USA Adele E. Goldberg Max Planck Institute for Evolutionary
Princeton University, USA Anthropology, Germany
William Croft
University of New Mexico, USA Seizi Iwata Arnold M. Zwicky
Kansai University, Japan Stanford University, USA
Charles J. Fillmore†
International Computer Science Paul Kay
Institute, Berkeley, USA University of California, Berkeley,
USA
Volume 17
Argument Structure in Usage-Based Construction Grammar. Experimental and
corpus-based perspectives
by Florent Perek
Argument Structure
in Usage-Based
Construction Grammar
Experimental and corpus-based perspectives
Florent Perek
University of Basel
doi 10.1075/cal.17
Cataloging-in-Publication Data available from Library of Congress:
lccn 2014048033 (print) / 2014050036 (e-book)
isbn 978 90 272 0439 4 (Hb)
isbn 978 90 272 6875 4 (e-book)
Acknowledgments ix
chapter 1
Introduction1
1.1 Overview: Argument realizationâ•… 1
1.2 Usage-based linguisticsâ•… 6
1.3 Structure of the bookâ•… 10
1.3.1 Part I: Verbsâ•… 10
1.3.2 Part II: Constructionsâ•… 11
1.3.3 Part III: Alternationsâ•… 11
chapter 2
Usage-based perspectives on verb valency 15
2.1 The verb in argument realizationâ•… 15
2.1.1 Projectionist approaches and their limitsâ•… 15
2.1.2 Constructional approachesâ•… 23
2.2 The division of labor between verbs and constructionsâ•… 27
2.2.1 The problem of mapping form to meaningâ•… 28
2.2.2 The need for richer lexical knowledgeâ•… 33
2.3 A usage-based account of verb valencyâ•… 37
2.4 Summaryâ•…42
chapter 3
Empirical evidence for usage-based valency 45
3.1 The hypothesis and its predictionsâ•… 45
3.1.1 The usage-based valency hypothesisâ•… 45
3.1.2 Existing evidence: Verb biases in language comprehensionâ•… 47
3.2 Assessing cognitive accessibilityâ•… 49
3.2.1 Why these verbs?â•… 50
3.2.2 Designâ•… 53
3.2.3 Stimuliâ•… 54
3.2.4 Participants and procedureâ•… 56
3.2.5 Resultsâ•… 57
vi Argument Structure in Usage-Based Construction Grammar
chapter 4
The usage basis of constructional meaning 79
4.1 The lexical origin of constructional meaningâ•… 79
4.1.1 Constructions and constructional meaningâ•… 79
4.1.2 Distributional biases and their significanceâ•… 80
4.1.3 Experimental evidenceâ•… 83
4.1.4 Evidence from corpus linguistics: Collostructional analysisâ•… 84
4.1.5 Summary: The usage basis of constructional meaningâ•… 89
4.2 Problems with the lexical basis of constructionsâ•… 90
4.2.1 The conative constructionâ•… 90
4.2.2 The semantics of the conative constructionâ•… 94
4.2.3 The conative construction in useâ•… 98
4.3 Conclusionâ•…102
chapter 5
The importance of local generalizations 105
5.1 Low-level schemasâ•… 105
5.1.1 Varying degrees of schematicityâ•… 105
5.1.2 The status of low-level schemasâ•… 108
5.1.3 Conclusion: The importance of local generalizationsâ•… 111
5.2 Low-level schemas in the conative constructionâ•… 111
5.2.1 Verb-class-specific constructionsâ•… 111
5.2.2 Verb-class-specific collexeme analysisâ•… 115
5.2.3 Verbs of ingestionâ•… 124
5.2.4 Verbs of cuttingâ•… 127
5.2.5 Verbs of pullingâ•… 129
5.2.6 Verbs of hittingâ•… 134
5.3 Summary and conclusionâ•… 139
chapter 6
Alternations as units of linguistic knowledge 145
6.1 Argument structure alternationsâ•… 145
6.2 Alternations in construction grammarâ•… 147
Table of contents vii
Conclusion
chapter 8
Summary and evaluation 211
8.1 Usage-based perspectives on argument realizationâ•… 211
8.2 Theoretical and methodological implicationsâ•… 216
References219
Appendix235
Constructions index 239
Name index 241
Subject index 245
Acknowledgments
Introduction
All languages provide ways to talk about events and their participants; this function
is typically assumed in great part by verbs. It is precisely for this reason that, more
so than other content words, verbs are rarely uttered in isolation but are usually
accompanied by certain other words, called the arguments of the verb. This book
is concerned with the topic of argument realization, i.e., that part of the grammar
that determines how participants to verbal events are expressed in the clause.1
For example, the verb kill is typically used with reference to at least two argu-
ments, a killer and a victim. Speakers of English must know that each of these
arguments is identified by a specific position in the clause: in the canonical word
order, the killer is realized in the pre-verbal position (called the subject in tra-
ditional grammar), while the victim is realized in the post-verbal position (the
direct object), as in Brutus killed Caesar. This knowledge constitutes the argument
structure of the verb kill. By contrast, there are other two-place predicates that
behave differently. For instance, the verb stare is commonly used in conjunction
with two arguments, an observer and a target. The latter of these arguments is not
realized as a direct object but as a prepositional phrase headed by at, as shown by
(1a) vs. (1b) below.
(1) a. *He stared me.
b. He stared at me.
The observations made so far may give the impression that argument structure
trivially consists of knowledge tied to individual verbs. However, the following
examples from the Corpus of Contemporary American English (COCA; Davies,
2008) do not fit well with this view.
1. It should be duly noted at the outset that the terms “argument realization” and “argument
structure” are not restricted to verbs, but can a priori apply to any kind of word that can be seen
as “governing” other elements in the sentence. It is, however, verbs that have received the most
attention in the literature on argument realization, probably because they can be seen, as Levin
& Rappaport Hovav (2005, p.â•›33) put it, as “the prototypical predicators, that is, argument-taking
words”. Some studies also apply the related term of “valency” to other parts of speech, such as
nouns, adjectives, and prepositions (Herbst, 1983; Herbst & Schüller, 2008).
2 Argument Structure in Usage-Based Construction Grammar
These examples strike us as rather creative. What they have in common is that they
feature an unusual use of each of the verbs. As previously mentioned, stare usually
takes two arguments and describes an event of looking, but its use in (2a) with a
third argument (into immobility) also conveys the causation of a change of state.
Similarly, the verbs cough, whistle, and sneeze typically refer to actions involving a
single individual, and we certainly do not want to claim that they inherently convey
the idea that these actions may result in the motion of some external entity (smoke/
her/blue pollen), as in (2b–d).
In the face of such examples, the idea that argument structure is primarily
knowledge about verbs loses some of its appeal. Speakers are able to use verbs in new,
creative ways to fit their communicative needs, which points to broader principles of
argument realization. As a matter of fact, such principles are necessary to explain cer-
tain facts about children’s early use of language, in particular their overgeneralization
errors (Baker, 1979), as exemplified by (3a) and (3b) below (from Bowerman, 1982a).
(3) a. I’m just gonna fall this on her.
b. I disappeared a bear in the back of the car.
These two naturally occurring sentences uttered by preschoolers involve the combina-
tion of a verb (fall and disappear) with an argument structure that is not acceptable in
adult usage. Since the children cannot have heard these odd combinations from their
caregivers, they must have applied a generalization beyond what is normally acceptable.
They would not be able to do so if they had only learned facts about individual verbs.
It is sometimes difficult to decide on purely intuitive grounds whether a given
use of an argument structure can be attributed to the main verb alone, or whether
it is better captured as a creative application of a generalization. Such problems will
be the central concern of Part I of this book. For example, many transitive verbs
in English can be used with an indirect object to refer to the intended recipient of
the result of the action described by the verb, as exemplified by (4) and (5) below
(also from COCA).
(4) a. Ask the butcher to grind you a little pork.
b. Jerry lit us a candle from the emergency kit.
c. I’ll just chop you a bit of wood.
(5) a. His mother cooked us an Indian-style meal.
b. Can I buy him a cookie?
c. I’ll write you a letter every day.
Chapter 1.╇ Introduction 3
It is unproblematic to claim that the sentences under (4) are creative uses, since
the verbs grind, light, and chop do not in themselves make reference to an intended
recipient or beneficiary (you in [4a] and [4c], us in [4b]), and there is nothing in
the meaning of these verbs that could predispose them to occur with one. The
sentences under (5) could be analyzed along similar lines, although the term “cre-
ative” seems less apt. Certainly, cook, buy, and write do not necessarily involve an
intended recipient of, respectively, the prepared food, the purchased goods, or the
written material. However, preparing meals for other people, purchasing items for
someone, or writing some text (a letter, a prescription, a recommendation, etc.)
intended for somebody else, are common occurrences, and the corresponding
formulations “cook/buy/write someone something” will most likely be familiar to
speakers of English. Tallying such uses with those under (4) fails to capture this
perceived conventionality. A similar problem arises more generally with optional
arguments. For example, the verbs rent and sell arguably presuppose a recipient,
which can be specified as an indirect object, but can also be omitted, in which case
the recipient argument receives a generic or indefinite interpretation:
(6) a. She rented (them) the apartment.
b. They sold (us) the house.
If there is some mechanism that allows speakers of English to use transitive verbs
with an additional recipient argument, and since the transitive use of rent and sell
is possible and does not produce a sense of “incompleteness”, why would it not be
assumed that rent and sell are essentially transitive and that their occurrence with
a recipient argument is licensed by the same mechanism that produces (4) and (5)?
Linguists have traditionally relied on their intuitions to decide such matters. I will
suggest an alternative approach, based on linguistic usage (cf. Chapters 2 and 3).
The examples discussed so far also illustrate another important property of
argument structure, i.e., that it produces meaning. Verbs can typically occur with
more than one argument structure, with systematic variations in meaning. For
instance, using a transitive verb with an indirect object, as in (4) and (5), introduces
the idea of an intended transfer. The following examples from Goldberg (1995,
p.â•›11) illustrate how different argument structures introduce semantic differences
between uses of the same verb, kick:
(7) a. The horse kicks.
b. Pat kicked the ball.
c. Pat kicked at the football.
d. Pat kicked Bob the ball.
e. Pat kicked the football into the stadium.
f. Pat kicked Bob black and blue.
4 Argument Structure in Usage-Based Construction Grammar
These sentences share the notion that the agent performs some gesture, namely
a forward motion of the foot, but they differ in many other respects. In (7a), the
horse acts alone, in (7b) and (7c), Pat’s action is targeted at a specific object, and in
(7d), it causes Bob to receive the ball. In (7e), Pat’s action on the football causes it to
move along a defined trajectory, and in (7f), Pat’s action on Bob imparts a change
of state on the latter.
Trivially, sentences with a different number of arguments are likely to differ
in meaning, to the extent that the semantic interpretation must be adapted to
accommodate additional arguments. Yet, even pairs of sentences with the same
number and type of arguments also exhibit meaning differences, as seen in the
contrast between (7b) and (7c). In the most likely scenario, (7b) refers to a situa-
tion in which John kicks and makes contact with the ball, which causes it to move.
However, (7c) does not necessarily imply that the ball was set in motion, neither
that contact was made; rather, it merely describes an attempt by John to cause the
ball to move (cf. Levin, 1993, pp.â•›64–65).
While descriptive accounts of the semantic aspects of argument structure
abound (Jackendoff, 1983; Dixon, 1991; Rappaport Hovav & Levin, 1998; Pinker,
1989; Goldberg, 1995, inter alia), linguists have more rarely addressed the ques-
tion of how argument structure comes to be associated with meaning in the first
place, and what determines this meaning. Part II will be concerned with this ques-
tion. One popular view is that argument structures derive their meaning from the
verbs that frequently occur in them (cf. Goldberg, Casenhiser, & Sethuraman, 2004;
Stefanowitsch & Gries, 2003). For instance, the notion of transfer conveyed by the
ditransitive argument structure is taken to originate in its frequent occurrence with
give and other verbs of giving. While this view seems adequate for argument struc-
tures that have a relatively concrete and stable meaning, it is less apt at capturing
cases where the semantic contribution is more abstract and variable (cf. Chapter 4).
As already mentioned, the insertion of at after the transitive verb kick usually pro-
duces an interpretation where contact with the direct object referent is not made.
For one thing, it is not clear what verb(s) the meaning ‘lack of contact’ could stem
from. Also, there are similar pairs of sentences with other verbs that do not display
the same semantic contrast, as exemplified with pull in (8a) vs. (8b) below. Both sen-
tences certainly entail that Bill made contact with the lever; the difference between
them lies in that (8a) but not (8b) entails that the lever was successfully moved.
(8) a. Bill pulled the lever.
b. Bill pulled at the lever.
This issue relates to the question of what level of generalization best captures argu-
ment structure: how can the structures used in (7c) and (8b) be considered the same if
they do not constitute a consistent semantic generalization? Can a single overarching
Chapter 1.╇ Introduction 5
The fact that many of such sentence pairs involve little variation in meaning (if
any) might lead us to consider the two patterns as grammatical variants, and sug-
gest that one could be derived from the other (and possibly vice versa), or at
least related to it in some way. On the other hand, each variant is associated with
particular constraints that restricts its use (cf. Erteschik-Shir, 1979; Thompson,
1990; Bresnan, Cueni, Nikitina, & Baayen, 2007). For instance, the double-object
variant is unacceptable with lexical recipients and pronominal themes (at least
in American English), ruling out sentences such as *He gave Larry it. Hence, the
two variants are not always interchangeable, which suggests that they should be
regarded as separate patterns. Similar observations can be shown to hold to varying
extents for many other pairs of argument structures (see Levin, 1993 for a number
of potential examples), which could be analyzed along the same lines.
Theories of argument structure have favored either one or the other posi-
tion, although there is arguably some merit in both: the former position explicitly
captures the relatedness between variants (e.g., Jackendoff, 1975; Pinker, 1989),
while the latter emphasizes the specific function of each (e.g., Goldberg, 1995,
2002; Michaelis & Ruppenhofer, 2001). This begs the question of whether these
two seemingly opposite views could in fact be reconciled (cf. Chapter 6). A more
interesting question, however, is whether both are needed to account for the lin-
guistic behavior of speakers (cf. Chapter 7). For instance, to what extent does the
occurrence of a verb in one of two related variants trigger the expectation that it
can also be used in the other variant, and how adequately do the two perspectives
capture that behavior?
This book examines argument realization from the perspective of a usage-
based approach to grammar, i.e., under the assumption that crucial aspects of
grammatical organization are tied to the frequency with which particular words
and syntactic structures are used. More particularly, it addresses the issues sketched
above by appealing to usage-based explanations. In the next section, I briefly in-
troduce the theoretical framework of usage-based linguistics and its basic tenets,
and I motivate its use for the study of argument realization.
6 Argument Structure in Usage-Based Construction Grammar
Usage-based theory takes the view that the cognitive representation of language
emerges through, and is shaped by, language use (Langacker, 1987, 2000; Hopper,
1987; Barlow & Kemmer, 2000; Bybee, 2006, 2010, 2013; Bybee & Hopper, 2001;
Bybee & McClelland, 2005). Hence, usage-based approaches reject the strict sepa-
ration of grammar (or competence) and usage (or performance) typical of genera-
tive approaches, and rather than construing grammar as a mere repository to be
accessed in language use, they take the view that grammar is itself the product of
usage. In Bybee’s (2006, p.â•›1) words, “grammar [is] the cognitive organization of
one’s experience with language”. Methodologically, it follows from this approach
that one can gain insights into the language system by analyzing usage data.
In usage-based approaches, grammar is commonly seen as a vast inventory
of symbolic conventions that are extracted from full-fledged utterances through
a gradual process of schematization, retaining the syntactic and semantic com-
monalities across different usage events. In line with this account of the emergence
of grammar, another important point of departure from generative approaches is
the exposure of the so-called “rule/list fallacy” (Langacker, 1987, p.â•›29), whereby
linguistic units must either be produced by maximally abstract rules of grammar,
or be listed in a lexicon containing all the irregularities. Instead, in a usage-based
approach, grammatical patterns may be defined at any level of abstraction. Specific
instances and the structures abstracted from them can be stored simultaneously; in
other words, lower-level units need not be discarded once a higher-level generaliza-
tion over these units has been made. In line with current thinking on categorization
in cognitive psychology (Medin & Schaffer, 1978; Nosofsky, 1986, 1988; Nosofsky,
Pothos, & Wills, 2011, inter alia), some recent versions of the usage-based approach
(notably Pierrehumbert’s [2001] and Bybee’s [2006, 2010, 2013]) propose that
the cognitive representation of language essentially consists of a large number of
exemplars stored in all their details in the course of language use. In these models,
abstract structures tend to be viewed as an emergent property of exemplar storage
rather than as explicit rules and schemas.
While the core tenets of usage-based linguistics can in theory be under-
stood independently of particular grammatical frameworks, they are in practice
susceptible to displaying varying degrees of compatibility with specific models.
Functionalist approaches (e.g., Givón, 1984, 1990; Dik, 1989; Halliday, 1994; Van
Valin & LaPolla, 1997) will generally tend to have more affinity with a usage-based
conception of grammar than generative frameworks (e.g., Chomsky, 1965, 1981,
1995; Borer, 2003; Ramchand, 2008; Randall, 2010). Due to their theoretical kin-
ship with usage-based linguistics and some shared assumptions about the nature
of language and its relation to the rest of cognition, cognitive-linguistic approaches
Chapter 1.╇ Introduction 7
when the conditional probability between two successive words is particularly high
(see also Gregory, Raymond, Bell, Fosler-Lussier, & Jurafsky, 1999). Hilpert (2008)
finds a corresponding effect of co-occurrence frequency on speech perception:
when hearing words rendered phonologically ambiguous by manipulation (e.g., cry
altered to sound more like try), speakers perceive the form that is the most likely
given the syntactic context. At the clausal level, there is a wealth of psycholinguistic
evidence that sentences tend to be processed more easily when individual verbs
are used with complements that are statistically more likely (Garnsey, Pearlmutter,
Myers, & Lotocky, 1997; Wiechmann, 2008; Wilson & Garnsey, 2009). Similar ef-
fects are reported in other domains of syntax, such as noun phrase conjunction
(Desmet & Gibson, 2003) and relative clause attachment (Desmet, De Baecke,
Drieghe, Brysbaert, & Vonk, 2006; Reali & Christiansen, 2007). Gahl & Garnsey
(2004) also find that verbs taking either a direct object (e.g., I heard the story yes-
terday) or a sentence complement (e.g., I heard the story was true) are pronounced
faster and with more final /t,d/ deletions when the post-verbal constituent matches
the probabilistic syntactic preference of the verb as measured by usage frequency.
Tily et al. (2009) report a similar finding with dative verbs alternating between
double-object (e.g., that gives them the full right to test the kids) and prepositional
constructions (e.g., I haven’t given much thought to it), for which they find that
the post-verbal complements are produced faster and with fewer disfluencies
(filled pauses, hesitations, repairs, etc.) when they match the construction that is
more likely according to a range of frequency-based factors (see also Kuperman &
Bresnan, 2012). Finally, frequent syntactic patterns also tend to be more resistant to
change (Bybee & Thompson, 1997). For example, Ogura (1993) observes that the
spread of do-support in questions (e.g., do you like…? instead of like you…?) and
negative sentences (e.g., you do not like… instead of you like not…) was resisted the
longest by high-frequency verbs such as say, think, and know. By the same token,
Bybee (2010) argues that the syntactic properties of auxiliary verbs (be, have, and
the modal auxiliaries can, must, may, will, etc.) in Present-Day English are largely
explained by the high frequency of occurrence of these verbs in the older question
and negation constructions without do-support (viz. “can/must/will/… subject
verb …?” and “subject can/must/will/… not verb”).
All of these findings, and many more, are not easily captured by an approach
that maintains a unidirectional relation from grammar to usage. By contrast, such
effects are naturally predicted by a usage-based account, and actually provide evi-
dence for one important corollary of the usage-based hypothesis, namely that differ-
ences in frequency should result in differences in linguistic structure. The impact of
frequency on linguistic representations is often discussed in terms of entrenchment,
which Langacker (2000, p.â•›3) defines as the degree to which “a highly complex event
can coalesce into a well-rehearsed routine that is easily elicited and reliably executed”.
Chapter 1.╇ Introduction 9
grammar is ultimately tied to usage, and, presumably, no less may be said about
the grammar of verbs. In fact, some of the findings mentioned above do point
to the direct relevance of usage to argument realization (e.g., Gahl & Garnsey,
2004; Wiechmann, 2008; Wilson & Garnsey, 2009; Tily et al., 2009; Kuperman &
Bresnan, 2012). Yet, many models still rely on traditional assumptions about the
nature of syntax and lexis inherited from research prior to the usage-based turn,
many of which have not been thoroughly examined from a usage-based perspec-
tive. This book seeks to mend this gap, by investigating usage-based principles of
the organization of argument realization.
The question of the usage basis of argument realization will be addressed at three
levels of organization, from the more specific to the more general: first verbs, then
grammatical constructions, and finally cross-constructional generalizations (tra-
ditionally known as argument structure alternations).
All models of argument realization incorporate in one way or another the fact
that verbs make reference to a number of participants. However, many verbs (if
not most) can occur with more than one set of arguments (or valency pattern). As
discussed in Chapter 2, so-called projectionist approaches claim that the verb is
stored with several different argument structures, but in construction grammar,
such an assumption is unnecessary, since, in that framework, constructions may
select only a subset of the arguments of a verb for overt realization, or conversely,
may contribute arguments by themselves. This, however, results into some in-
determinacy as to how many arguments are contributed by the verb in a given
instance, which in many cases cannot be decided on intuitive grounds, especially
if it is assumed that a verb should only have one lexical entry. I suggest, following
a proposal by Langacker (2009), that verbs can be stored with a range of valency
patterns that is directly related to that verb’s usage, in that repeated use of a verb
in a construction leads to the conventionalization of the corresponding valency
pattern with that verb, which as a result comes to be associated directly with it,
and not arrived at via combination with a construction.
In Chapter 3, I test this hypothesis by comparing experimental results with
usage data. I report on an incremental reading comprehension experiment de-
signed to determine the relative cognitive salience of two valency patterns of three
commercial transaction verbs: buy, pay and sell. I then compare the results of this
Chapter 1.╇ Introduction 11
experiment to usage data extracted from corpora. As it turns out, the frequency
of a verb in a specific syntactic environment correlates with the relative cognitive
salience of the corresponding valency pattern for that verb (as measured by read-
ing times), in that more frequent valency patterns are processed more easily. These
findings provide evidence for the usage basis of verb valency.
The notion of alternations in linguistics refers to the possibility for the same unit
to receive different formal realizations. In the grammar of verbs, pairs of con-
structions corresponding to different realizations of a common set of arguments
have also been characterized as syntactic alternations, e.g., the dative alternation,
12 Argument Structure in Usage-Based Construction Grammar
pairing the double-object variant (John gave a book to Mary) with the to-dative
variant (John gave a book to Mary). In construction grammar, the variants of an
alternation are seen as independent constructions with their own properties (cf.
Michaelis & Ruppenhofer, 2001; Goldberg, 2002), but the relation between them is
usually not considered part of the grammar. In Chapter 6, I question whether this
view really provides a good account of what speakers know about their language.
Adhering to Cappelle’s (2006) allostructions model, I suggest that speakers store
generalizations of a common constructional meaning over formally distinct argu-
ment structure constructions, and I report experimental evidence for this proposal.
I also suggest that such alternation-based generalizations play a role in argument
structure productivity.
In Chapter 7, I report on an experimental study testing for productivity asymme-
tries in the dative alternation and in the locative alternation (spray paint on the wall
vs. spray the wall with paint). The experiment examines whether the extent to which
speakers productively use a novel verb in the other variant of an alternation varies
according to the variant with which this verb was previously presented to them. I find
that there is indeed a productivity asymmetry in the dative alternation, whereby
subjects are more likely to use the verb productively in the unattested variant when
the novel verb was first presented in the double-object construction than when it
was presented in the to-dative construction. In the locative alternation, no asym-
metry is found. I argue that the productive behavior of speakers can be explained by
patterns of usage. Drawing on corpus data, I show that many more dative verbs are
attested in the to-dative variant only than in both constructions; conversely, almost
all of the verbs attested in the double-object variant are also attested in the to-dative
variant. In other words, the assumption that a given verb alternates is much more
likely if that verb has been observed in the double-object variant than in the to-dative
variant. Such patterns of type frequency are not found for the locative alternation,
which explains its lack of asymmetry. I suggest that speakers have internalized these
patterns of usage, and that their knowledge of grammar includes information about
relations between constructions involved in an alternation.
In sum, the findings reported in this book show that the cognitive representa-
tions involved in argument realization are largely shaped by usage, at all three levels
of organization. In Chapter 8, I elaborate on some theoretical and methodological
implications of these findings.
part i
Verbs
chapter 2
In this chapter, I take verbs as the starting point of my study of argument realiza-
tion. The chapter is structured as follows. I start with a general discussion of the
role of the verb in contemporary theories of argument realization, contrasting
projectionist and constructional approaches, and I show that there are good rea-
sons to favor the latter kind. I then argue that constructional approaches have not
entirely resolved the question of the mapping of form to meaning in the structure
of clauses, in that it is often unclear what the contribution of the verb exactly is. To
address this issue, I argue for a usage-based approach to verb valency, according to
which the contribution of verbs to clauses depends on their prior usage.
It is precisely by virtue of their relational character that verbs are not used in
isolation but accompanied by a number of arguments.2 In cognitive terms, verb
1. Original quote: “Le noeud verbal […] exprime tout un petit drame. Comme un drame en
effet, il comporte obligatoirement un procès, et le plus souvent des acteurs et des circonstances.”
2. The only exception is a very small class of verbs describing meteorological phenomena like
rain and snow, which can be conceptualized by themselves without making reference to particular
entities. In English, such verbs are canonically accompanied by at least a ‘dummy’ subject pronoun
it (e.g., it rained for days), which is not referential and therefore does not designate a participant
16 Argument Structure in Usage-Based Construction Grammar
of the event. In languages that, contrary to English, allow declarative sentences without a subject,
finite forms of meteorological verbs stand on their own (cf. Italian piove ‘it’s raining’).
3. This frame-based analysis of argue was taken from the database FrameNet (https://framenet.
icsi.berkeley.edu), the main lexicographic implementation of Frame Semantics (cf. Section 5.2.2).
Chapter 2.╇ Usage-based perspectives on verb valency 17
A purely lexicalist model of argument realization is also at odds with the facts
of language acquisition. A lexicalist view of argument structure entails conser-
vative means of acquisition: under this account, children learn the argument
structures properties of a verb by observing in which syntactic environment(s)
the verb is used. Consequently, they could not be able to produce combinations
that they have not previously heard, but there is empirical evidence that they in
fact do: as mentioned in Chapter 1, examples of so-called overgeneralization er-
rors are well-known and abundantly documented by language acquisition studies
(Bowerman, 1982a, 1982b, 1988; Gropen, Pinker, Hollander, Goldberg, & Wilson,
1989; Mazurkewich & White, 1984; Pinker, Lebeaux, & Frost, 1987). For example,
(1a) below shows the verb say used with a recipient realized as an indirect object
instead of the conventional to-phrase, and in (1b), the verb cover is mistakenly used
with a direct object theme and a locative prepositional phrase, instead of a direct
object goal/location and a with-phrase theme (i.e., cover me/myself with a screen).
(1) a. Don’t say me that or you’ll make me cry. (Bowerman, 1988, p.â•›79)
b. I’m gonna cover a screen over me. (ibid., p.â•›80)
Hence, children are able to use verbs in argument structures in which they could
not have been previously heard, which points to the existence of verb-general
mechanisms that assign argument structure information to verbs.
18 Argument Structure in Usage-Based Construction Grammar
the meaning of eat as an event consisting of the act predicate, which takes two
arguments (the thing nodes), and an optional manner component lexically speci-
fied by the verb. The empty square brackets under the thing nodes indicate open
argument positions. This semantic representation can be paraphrased as “a thing
acts on another thing in an eating manner”. Of course, this overlooks many facts
that we know about eating: that it is usually performed by a living being, that it
involves that some solid substance goes inside this being to be processed, etc. The
lexical representation contains only those aspects of meaning that are sensitive to
argument realization. Hence, as far as grammar is concerned, the linguistic mean-
ing of eat does not fundamentally differ from other verbs such as drink, cook or
kick, in that their lexical semantic structure are also based on the act predicate,
albeit with a different manner component in each case.
eat: EVENT
Figure 2.1╇ Semantic structure of the verb eat (from Pinker, 1989, p.╛206).
EVENT EVENT
SUBJECT OBJECT
Figure 2.2╇ Linking rules for the act predicate (from Pinker, 1989, p.╛206).
the argument structure of any verb based on the act predicate, without the need to
stipulate the mapping from semantics to syntax in each individual case (contrary
to the lexicalist approaches mentioned earlier). Importantly, the same would hold
for any verb of that kind that would enter the lexicon.
eat: EVENT
SUBJECT OBJECT
Figure 2.3╇ Argument structure of the verb eat following Pinker’s (1989) model.
undergo a particular change of state”. These event structure are canonically real-
ized respectively as a transitive clause, e.g., Phil swept the floor, and as a resultative
construction, e.g., Phil swept the floor clean.
There are, however, cases in which differences in argument realization do
not evidently correlate with differences in meaning pertaining to the whole event
structure. For example, the post-verbal arguments of verbs entering into the da-
tive alternation, such as give and tell, can be encoded either in two noun phrases
or in a noun phrase and a prepositional phrase headed by to, with no substantial
difference in propositional meaning:
(6) a. Mary gave John a book.
b. Mary gave a book to John.
(7) a. John told Mary a joke.
b. John told a joke to Mary.
Either use of give and tell refers to the transfer of an object or a message to a
recipient, and can arguably be used to describe the same range of situations (cf.
Rappaport Hovav & Levin, 2008 for further arguments against a multiple-sense
approach to dative alternation verbs). There is therefore little reason to assume
that give and tell have a different meaning when used in each variant of the dative
alternation, other than the need to preserve a one-to-one mapping between lexical
semantics and argument realization, along the lines of a projectionist approach.
Hence, such an account appears opportunistic, and, as Goldberg (1995, pp.â•›10–12)
observes, ultimately circular. It makes the assumption that give and tell must have
two distinct senses because they are used in these two argument structures, while
simultaneously claiming that it is precisely because they have these two senses that
they can be used in these argument structures.
Another domain in which the multiple-sense approach clearly shows its limits
is regarding unconventional uses of verbs, as exemplified by the following sen-
tences from the COCA (Davies, 2008), repeated from Chapter 1:
(8) a. Chess coughed smoke out of his lungs.
b. [H]er nose was so bloodied that the ref whistled her off the floor.
c. Navin sneezed blue pollen onto his shirt.
(9) a. Ask the butcher to grind you a little pork.
b. Jerry lit us a candle from the emergency kit.
c. I’ll just chop you a bit of wood.
sentence describes in each case the motion of some entity caused by the intransitive
action. In (9a–c), the transitive verbs grind, light, and chop are used with a third,
indirect object complement.
In a projectionist account, these uses would be considered instances of particu-
lar lexical entries with specific meanings corresponding to these syntactic frames:
cough, whistle, and sneeze would be taken to mean ‘cause something or someone
to move by coughing/whistling/sneezing’, and grind, light, and chop, ‘grind/light/
chop something with the intention of giving it to someone’. As argued by Goldberg
(1995, pp.â•›9–10), such highly specialized senses seem counter-intuitive, and it is
hard to imagine a language in which such meanings would be lexicalized by a
separate stem. Surely, no scholar working in a projectionist framework would claim
that speakers actually store a lexical entry for these verbs with the corresponding
meanings; more plausibly, such entries would be formed ad hoc to match the verb
with the syntactic frame it occurs in. Many projectionist accounts include some
mechanism (called lexical rules in Pinker’s [1989] model) that derive new lexical
entries (with the corresponding argument structure) from existing ones provided
they match certain conditions on semantic structure (cf. Bresnan, 1982; Briscoe &
Copestake, 1999; Müller, 2006; Pollard & Sag, 1984, inter alia); such mechanisms
are also relied on to account for the acquisition of argument structure. Be that as
it may, the very idea that a verb would only adopt a particular meaning when used
in a particular syntactic environment lacks credibility, and undermines a lexical
treatment of creative uses such as those listed in (8) and (9) (but see Müller &
Wechsler [2014] for a defense of the lexical rules approach, and Goldberg [2014],
Boas [2014], and the other papers in the same issue for a response). Hence, the
existence of such sentences weakens the assumption that argument realization
draws on lexical information alone (see also Goldberg, 2013).
The problems faced by projectionist approaches have led some researchers
to abandon the assumption that verbs are the sole bearer of argument realization
information. Instead, there is a growing consensus that many aspects of the form
and meaning of clauses are contributed by syntactic constructions themselves.
These so-called constructional approaches are presented in the next section.
several functionalist and even generative frameworks (e.g., Borer, 2003; Jackendoff,
2002; Van Valin, 2005). However, since the theoretical assumptions behind Goldberg’s
theory are more in line with the usage-based approach to argument realization that I
adopt in this book, I will mainly focus on Goldberg’s original model.
The central idea of Goldberg’s theory is that argument structures are not pro-
jected by verbs; instead, they are stored as independent form-meaning pairs, or
constructions. Argument structure constructions pair an abstract event descrip-
tion containing an array of arguments, with a specification of the morphosyntactic
encoding of these arguments. For example, the ditransitive construction refers to
a three-argument scene of transfer, paraphrased as ‘X cause Y to receive Z’, and
specifies that X (the agent) is realized as the subject of the clause, Y (the recipient)
as the indirect (or first) object and Z (the patient) as the direct (or second) object.
This is diagrammed in Figure 2.4.
PRED
Figure 2.4╇ The ditransitive construction (adapted from Goldberg, 1995, p.╛50).
Since the argument roles of a construction are defined with reference to the meaning
of that construction, and the participant roles of a verb are defined with reference
to the frame-semantic meaning of the verb, the semantic coherence principle can
be more generally understood as requiring that the semantic frame evoked by the
verb be semantically compatible with the meaning of the construction.
Chapter 2.╇ Usage-based perspectives on verb valency 25
Goldberg’s stipulation that the participant roles of the verb must be construed as,
though not necessarily be, instances of the argument roles of the construction,
allows for some flexibility in determining semantic compatibility. In some cases,
the construal of a verbal meaning as an instance of a constructional meaning is
mediated by metaphorical mapping. For example, tell does not convey transfer of
possession in a strict sense: its occurrence in the ditransitive construction relies on
a metaphorical construal of communication as giving, whereby ideas and messages
are construed as objects that are given in an abstract sense to a hearer/recipient
(cf. Reddy’s [1979] “conduit metaphor”). Other cases rely on extensions of the
constructional meaning; for example, the occurrence of promise in the ditransitive
construction conveys not actual but future giving.
Since constructions are themselves meaningful, they can also contribute as-
pects of meaning to the clause that are not normally associated with the verb. Such
cases require that the meaning of the verb can be integrated into the meaning of
26 Argument Structure in Usage-Based Construction Grammar
PRED
Figure 2.6╇ The caused-motion construction (adapted from Goldberg, 1995, p.╛52).
In sentences such as (10), the meaning of sneeze describes the means whereby
motion is caused in the event meaning ‘X cause Y to go Z’ contributed by the
construction. Because the single argument of intransitive sneeze is the only acting
entity, and because its action (emission of a force) is susceptible to causing the
motion of another entity, the ‘sneezer’ role of sneeze is fused with the ‘agent’ role of
the caused-motion construction. However, the event meaning of the construction
contains two more argument roles (theme and path) that do not find matches in the
meaning of the verb: these arguments are contributed by the construction alone. In
sum, sneeze remains a one-argument predicate and conveys the same meaning as it
does elsewhere, and the missing aspects of meaning (event-level semantics and the
corresponding additional arguments) are contributed by the caused-motion con-
struction; this process is commonly discussed in terms of constructional coercion
(cf. Michaelis, 2005; Lauwers & Willems, 2011, and the papers in the same volume).
The fusion of the caused-motion construction with the verb sneeze is diagrammed
in Figure 2.7. As opposed to the case of give in the ditransitive construction illus-
trated in Figure 2.5, two argument roles of the caused-motion construction (theme
and path) are not fused with participant roles of the verb. The ‘means’ label on the
relation between the meaning of the construction and the meaning of the verb
captures the specific way in which they are combined.
Chapter 2.╇ Usage-based perspectives on verb valency 27
Figure 2.7╇ Fusion of the caused-motion construction with sneeze (adapted from
Goldberg, 1995, p.â•›54).
verb itself. I suggest that it is perfectly plausible to assume that a verb can be stored
with different conventional construals of the frame corresponding to different
sets of participant roles, and I show that such a view is actually in line with recent
evidence that a large amount of verb-specific information is needed in order to
account for the distribution of verbs and constructions.
1977). The use of sell in the ditransitive construction, as in (11) below, includes
the buyer within the windowing of attention (but not the money) and profiles the
entire transfer of goods from seller to buyer, thus construing the event as one of
caused possession.
(11) John sold Mary a computer.
If, however, sell is used in the transitive construction, only the seller and the goods,
respectively realized as subject and object as in (12) below, are comprised within
the windowing of attention, and the buyer is excluded from it. In that case, only
the onset of the transfer is profiled, and the event is construed as one in which the
seller does something with the goods, giving them away.
(12) John sold a computer.
How do we negotiate the mapping of form to meaning in such cases? The construal
that a construction imposes on a certain event is a semantic contribution in itself,
in the same capacity as in those cases where the constructional meaning is more
visible. Therefore, the question of which aspects of the meaning of sentences (11)
and (12) are respectively contributed by the verb and by the construction is a valid
one. Because constructions can coerce a verb meaning into a particular construal,
thereby adding arguments or leaving certain participant roles unexpressed, a con-
structional approach often allows the mapping of form to meaning to be resolved in
different ways (cf. Croft, 2012, p.â•›366 for a similar point). In particular, it is often not
clear what the valency of the verb is, i.e., which participant roles are contributed by
the verb itself. For (11), sell can be considered as a three-participant verb construing
the event of selling as a scene of transfer; in this case, the verb is inherently compat-
ible with the ditransitive construction, and the meaning of verb and construction
(as well as their arguments) overlap. Alternatively, since a two-participant construal
of selling is conceivable, it could be argued that this construal is the actual meaning
of the verb; in this case, the ditransitive construction would itself contribute the
three-participant construal and the additional indirect object argument. Similar
speculations on the meaning of sell also lead to two alternative analyses of the
mapping of form to meaning in (12): either as a two-participant verb inherently
compatible with the transitive construction, or as a three-�participant verb coerced
into a two-participant construal by the construction. Countless other examples
could be added to the case of sell to illustrate this point.
It might be argued that the existence of such issues precisely points to the prob-
lematic status of lexical entries, which perhaps should be abandoned altogether.
As previously discussed, verbs evoke semantic frames; it could be assumed that
the participant roles fused with the argument roles of constructions are directly
30 Argument Structure in Usage-Based Construction Grammar
drawn from frames, without an intervening level of lexical entries. This would,
however, have undesirable consequences. It would essentially reduce verbs to labels
for frames, which would fail to capture semantic differences between items evoking
the same frame, especially items from other word classes. The difference between
words evoking the same frame lies in what portion of this conceptual content
they profile, to use Langacker’s (1987, 2008) terminology. Langacker defines an
expression’s profile as the part of its “conceptual base” (i.e., its frame semantics)
that “stands out as the specific focus of attention within its immediate scope” (2008,
p.â•›66). Simply put, the profile is what the expression designates, as opposed to
what it merely presupposes. As Langacker (2008, p.â•›67) notes, many words differ
semantically not in terms of their conceptual import, but in terms of what part of
this content they profile, which is but one aspect of the construal they impose on
it. For example, the nouns hub, spoke and rim all have (in their relevant sense) a
wheel as their conceptual base, but they refer to (viz. profile) different parts of it.
Similarly, the agent nominalization seller differs from its verbal root sell in that it
profiles a particular frame element in the commercial transaction frame, while sell
profiles the relation between frame elements itself.
Profiling is lexically determined and can largely be arbitrary; at any rate, it is a
necessary part of a word’s meaning. In the case of verbs, it is possible that profiling
includes only a subset of the participants involved in the semantic frame, leaving
the others in the background. It could also be assumed that verbs simply profile
the entire frame, including all the frame elements. Note, however, that if all verbs
evoking the same frame receive the same profile, this could deprive us of a potential
way to capture semantic distinctions between them. In particular, it would entail
that all verbs evoking the same frame also give the same prominence to the differ-
ent frame elements, which is clearly not the case. For instance, both seek and forage
evoke the Seeking frame according to the FrameNet database, but seek clearly give
more importance to the sought entity than forage does, since this frame element
must be obligatorily expressed with the former but not with the latter, as seen by
the comparison of Examples (13) and (14).
(13) a. They sought food (in the woods).
b. *They sought in the woods.
(14) a. They foraged for food (in the woods).
b. They foraged in the woods.
Such differences are not expected if verbs are taken to equally profile all frame
elements. Moreover, it appears that for many frames, uses of a verb with all frame
elements expressed are actually very uncommon. For instance, assuming that the
commercial transaction frame contains at least four frame elements (buyer, seller,
Chapter 2.╇ Usage-based perspectives on verb valency 31
goods, and money), examination of corpus data from both British and American
English (cf. Section 3.3.2) reveals that uses of the verbs buy, pay, and sell (evoking
this frame) with all four frame elements (as exemplified by [15] below) account
for less than 5% (and in most cases less than 1%) of the distribution of these verbs.
(15) [W]hen you bought the property you could have sold it to them for fifty grand
more (CallFriend corpus)
4. There is a parallel to be drawn here with aspectual coercion (Michaelis, 2004, 2005), which
concerns how the temporal unfolding of the event described by a verb is altered in different uses.
Aspectual coercion can in many cases be described as variation in profiling of the temporal scale
of the event; it is for instance how Langacker (2008, 5.2) describes a change from perfective (viz.
bounded in time) to imperfective (viz. unbounded) aspect (see also Croft, 2012). Similarly, we
could call valency coercion the phenomenon whereby a different number of participants to the
event are included in a verb’s profile.
32 Argument Structure in Usage-Based Construction Grammar
a one-participant sense that does not involve caused motion, which is contributed
by the construction. However, it is not clear to what extent the same argument
applies to possible verb entries corresponding not to different conceptual content
but merely to different construals of the same frame. Intuitively, it does not seem
implausible that either construal of sell, as transferring possession or as losing pos-
session, is associated with the verb itself, since both are arguably regular ways to
conceive of selling events. For the same reason, it is not clear why one (and which
one) should be the ‘basic’ sense of the verb and the other derived by constructional
coercion. Langacker (2009, pp.â•›251–252) makes a similar observation, taking the
verb kick as example:
Suppose it is claimed that positing a caused-motion sense for kick would itself
be implausible. I would argue that such a claim is gratuitous, since construction-
related meanings (i.e., aspects of meaning pertaining to profiling, argument struc-
ture, etc.) cannot be ascertained independently of occurrence in constructions.
The kicking scenario does not come with any pre-assigned category or grammati-
cal frame. It lends itself to nominal construal (e.g. kicker) and to simple transitive
use, but since kicking often results in motion on the part of the object kicked, it
also lends itself naturally to the caused-motion construction.
In sum, while I generally adhere to Goldberg’s (1995) arguments against the pro-
liferation of construction-related lexical entries, I do not consider (in line with
Langacker) that the canonical situation is necessarily one where the verb is associated
with a single lexical entry per semantic frame and where the whole distribution of
Chapter 2.╇ Usage-based perspectives on verb valency 33
this lexical unit is arrived at by combining it with various constructions, which would
be the logical consequence. Another of Goldberg’s (1995, pp.â•›12–13) arguments for
limiting the number of lexical entries is “semantic parsimony”, i.e., that it is prefer-
able to ascribe recurring aspects of meaning to a construction rather than to assume
that they are stored redundantly in multiple lexical entries. However, as pointed out
by Croft (2003) and Langacker (2009), parsimony is no criterion for psychological
adequacy; as Langacker puts it, “while it is certainly more parsimonious to posit one
sense rather than several, our objective is to model linguistic knowledge correctly,
not to write the shortest grammar that generates all the grammatical forms” (p.â•›251).
Ultimately, the question of whether verbs are stored with multiple lexical en-
tries is part of the storage vs. computation debate. Having several verb entries that
are readily available for use in particular syntactic frames (i.e., without relying on
constructional coercion) constitutes a processing advantage, whereas positing a
single entry means that all uses departing from it are arrived at through combi-
nation with constructions, and therefore assumes more reliance on composition
processes. Parsimony is an explicit criterion in generative linguistics, aiming at
writing the smallest grammar possible, but it is irrelevant in cognitively-oriented
approaches, where both rules and specific instantiations of these rules (in this case,
construction-congruent lexical entries) can be redundantly stored. If psychological
adequacy is what cognitive linguists should aim at, then storage parsimony should
be the exception rather than the rule, as there is ample evidence that language
users rely to a large extent on prefabricated stored assemblies, as opposed to on-
line processing of creatively combined elements (cf. Ellis, 1996; Erman & Warren,
2000; Bannard & Matthews, 2008; Arnon & Snider, 2010).
In sum, I concur with Langacker (2009, p.â•›255) that “just a single verbal sense
should not be accepted by default as the canonical situation”. While we should
certainly not posit a new leical entry each time a verb is witnessed in a different
construction, it seems reasonable to hypothesize that verbs can be conventionally
associated with more than one construal of their semantic frame. As a matter of
fact, this position is consonant with recent research suggesting that a lot of in-
formation must be stored at the level of verbs if we are to fully account for their
syntactic behavior. I discuss this research in the next section.
Verbs can be restrictive as to the semantic type of result phrase they accept; for
example, paint in (16) seems to take only result phrases denoting a color, even
though other resulting properties of the patient would make sense. Moreover,
some verbs only allow the result state to be expressed as an adjective phrase, such
as hammer in (17), while others only allow a prepositional phrase, such as strangle
in (18). Finally, even verbs with a very similar meaning, like talk and speak in (19),
may differ in their behavior with respect to resultative uses.
Boas criticizes Goldberg’s (1995) account of such phenomena in terms of
maximally general constructions by pointing out that it cannot accurately predict
the full distribution of resultative sentences. Instead, he argues that a great deal of
verb-specific knowledge must be stored at the level of lexical entries (called “mini-
constructions” in Boas, 2008), including such information as semantic or syntactic
restrictions on the possible result phrases, typical collocations and world knowledge.
Along the same lines, Boas (2008, p. 132) notes that “while abstract constructions
may be sufficient for comprehension, for production we need to refer to more de-
tailed information in order to arrive at correct predictions about the distribution of a
verb’s arguments”. In sum, Boas’ account of resultatives emphasizes that while higher-
level schemas might be derived by noticing the similarities across many of these
individually learned “mini-constructions”, the actual distribution of resultatives can-
not be licensed by a highly abstract construction which would blur many important
combinatory details, and therefore is better captured at a more idiosyncratic level.
A similar point is made by Faulhaber (2011a, 2011b) on a yet larger scale. On
the basis of the corpus-derived Erlangen Valency Patternbank database (Herbst &
Uhrig, 2010) based on the Valency Dictionary of English (Herbst, Heath, Roe, &
Chapter 2.╇ Usage-based perspectives on verb valency 35
5. I am grateful to one of the anonymous reviewers for pointing out to me the convergence
between these strands of valency grammar (in particular Herbst’s work) and the present
approach.
36 Argument Structure in Usage-Based Construction Grammar
carriers. Against this backdrop, Herbst (2010, 2011) points out the complemen-
tarity between construction grammar and valency theory (see also Stefanowitsch,
2011), and combines the two approaches by recognizing the relevance of argument
structure constructions to account for creative uses of verbs, and by introducing
the idea of valency constructions. Valency constructions are described as form-
meaning pairs relating a participant pattern (viz. an array of participant roles) to a
valency pattern and its valency carrier; in other words, valency constructions are
essentially lexically-filled instantiations of argument structure constructions that
correspond to the established uses of a verb. In fact, Herbst (2011) sees Goldberg’s
(1995) argument structure constructions as generalizations over valency construc-
tions, a view which lines up with usage-based accounts of language acquisition
(Tomasello, 2003).6 In this combined model, Herbst argues that argument realiza-
tion is governed by the Valency Realization Principle, whereby a participant pattern
of a verb can only be realized as an instance of an argument structure construction
if there exists a corresponding valency construction for that verb (cf. Herbst, 2011,
p.â•›356). The only exception to this principle is when the combination of a verb with
an argument structure construction involves participant roles that are not part of the
verb’s valency slots, like in the creative uses listed in Example (8) (cf. p.â•›22). Under
this account, the grammar of verbs relies on a great deal of item-specific knowledge,
which is argued to be indispensable in order to capture the many idiosyncrasies of
individual verbs. Correspondingly, the role of argument structure constructions is
substantially reduced, as they are restricted to account for creative uses. Thus, in
essence, Herbst’s model can be seen as a projectionist (or more exactly lexicalist)
approach with a constructional “appendix”.
The principal merit of Herbst’s approach is to combine the best of both worlds,
by giving a more explicit role to verb-specific knowledge along with construc-
tions, which lines up with Goldberg’s (2006, p.â•›12) contention that “item-specific
knowledge exists alongside generalizations”. Another appealing aspect is that lexi-
cal units of argument realization are conventionalized form-meaning pairs instead
of decontextualized lexical entries, thereby providing continuity with argument
structure constructions (see also Goldberg, 2006; Boas, 2008). It is not clear, how-
ever, to what extent this model achieves cognitive adequacy, which, admittedly,
is not an objective that it explicitly pursues. In particular, the Valency Realization
Principle clearly appears to be too strong, since it fails to predict the existence of
6. Valency constructions are thus similar to Goldberg’s (1995, p.â•›51) “composite fused struc-
tures” resulting from the fusion of a verb and a construction, except that the former refers to
specific formal generalizations (i.e., they make reference to specific syntactic categories instead
of functional labels such as “object” or “locative”), and are explicitly defined as stored units.
Chapter 2.╇ Usage-based perspectives on verb valency 37
Langacker outlines a four-stage model detailing how the recurrent use of a verb in a
particular construction leads to the conventionalization of a new “sense” of the verb,
and illustrates this account with the case of the caused-motion use of kick. Figure 2.8
38 Argument Structure in Usage-Based Construction Grammar
represents the first stage of the model. Langacker starts with the assumption that
the verb kick is inherently associated with a construal corresponding to its transi-
tive use. This lexical entry is represented by the bottom left box in Figure 2.8;7 the
square angles of the box indicate that it is a conventional unit. The verb profiles
an exertion of force by the trajector (‘tr’) on the landmark (‘lm’), correspond-
ing to the heavy-lined parts of the diagram. The dashed arrow pointing from the
landmark marks that motion does not necessarily occur and is not profiled by the
verb. The caused-motion construction, represented by the top box in Figure 2.8,
is a constructional schema (viz. a complex symbolic assembly) that consists of two
components. The left component depicts a binary relation in which the trajector
exerts a force (marked by the double-arrow) on the landmark, and the landmark
is set in motion as a result. This component essentially corresponds to the verb
slot of the construction, and it is schematic for verbs of caused-motion like throw,
toss, lift, etc. (with their subject and direct object arguments). The right component
describes the path followed by the landmark of the binary relation, and basically
corresponds to the locative adverbial slot.
tr lm tr
T T′
tr lm tr lm
kick kick
7. All diagrams presented in this section are taken from Langacker (2009, pp.â•›252–254), with
slight adaptations.
Chapter 2.╇ Usage-based perspectives on verb valency 39
categorization (labeled T). In Figure 2.8, the lexical entry of kick instantiates the verb
slot of the construction, but there is a conflict in the semantic specifications of the
target and the standard of categorization. In the former, the motion of the landmark
is a mere implicature and is not profiled,8 whereas in the latter, motion necessar-
ily occurs and is profiled. Hence, the relation between the two is one of extension
rather than elaboration; following Langacker’s notation, this is indicated by a dashed
arrow pointing from the constructional slot to the verb. Langacker argues that in
such a case of partial recognition, the aspects of the standard that are missing in
the target are projected onto an extended target (labelled T′ in Figure 2.8). In other
words, the verb is apprehended as a verb of caused-motion, which is the Cognitive
Grammar way to represent the integration of constructional meaning with verbal
meaning. Because the extended target matches the constructional slot, a relation of
elaboration can be drawn between S and T′, marked by a plain arrow in Figure 2.8.
Initially, the apprehension of kick as a verb of caused-motion (represented by
the box surrounding S, T and T′), and the augmented target resulting from it, are
only transient and do not constitute conventional units, as marked by the rounded
angles. In other words, the “new” lexical entry of the verb (T′) only exists within
the context of the partial recognition involved by the categorization of transitive
kick (T) as a verb of caused-motion. In that sense, it is a genuinely creative use of
the verb, similar to the caused-motion use of sneeze discussed earlier. However, if
such a use is to recur, the recognition of kick as an extension of the constructional
slot becomes a conventional unit (this is the second phase of the model). At this
stage, there has been yet no change in the respective semantic contribution that
the verb and the construction make to the clause, but this use of the verb is now
familiar to speakers, and is no longer perceived as innovative. This conventionaliza-
tion (marked by a square box around the extended target and the categorization as
a whole in Figure 2.9) sets the necessary conditions for the next phase to happen,
namely the loss of analyzability, presented in Figure 2.9. Since there is a convention-
alized elaborative relation between the constructional slot and the “derived” lexical
sense (T′), the latter can serve as a direct target for categorization, thus eschewing
the need for an extension relation to be established between the constructional slot
and the original sense (T) first, as it is the case in innovative uses exemplified by
Figure 2.8. As a consequence, the original sense is not necessarily activated upon use
of the verb in the construction (this is represented by the dashed box in Figure 2.9);
the augmented target gradually loses its status as an extension from the original
meaning of the verb, and becomes the actual target of categorization.
8. Boas (2003, 7.3) presents a similar account of the conditions that license the occurrence of
air emission verbs such as cough and sneeze in the caused-motion construction. He also notes
the importance of contextual factors (see also Boas, 2011b).
40 Argument Structure in Usage-Based Construction Grammar
tr lm tr
T T′
tr lm tr lm
kick kick
The outcome of this process is presented in Figure 2.10. The original transitive
sense of kick is completely absent from this diagram. Instead, the target in the con-
ventionalized categorization corresponds to what previously was the augmented
target generated by the partial recognition of the original sense of kick as a caused-
motion verb in former usage events. In other words, a new sense of the verb has
emerged that is inherently compatible with the caused-motion construction (i.e.,
through elaboration, and not extension, of the verb slot in the construction). This
means that for the caused-motion use of kick, the mapping of form to meaning
has changed from a situation in which the construction conveys caused motion
and the verb conveys the means whereby caused motion occurs, to one in which
both verb and construction convey caused motion.
In this model, the difference between inherent compatibility and semantic
enrichment is a matter of gradience rather than a clear-cut distinction, and the
position of any composite structure on the continuum between either of these
extreme cases depends on how frequently the verb has been witnessed in the rel-
evant construction. Accordingly, the four stages of Langacker’s model should per-
haps not be understood as corresponding to discrete states, but rather as fuzzily
bounded regions on this continuum.
In Section 2.2.1, I discussed cases in which the occurrence of a verb in dif-
ferent constructions corresponds to different profilings of the semantic frame.
Langacker’s model predicts that these different profilings might well emerge as
different lexical entries of the verb. For example, assuming that sell initially has
Chapter 2.╇ Usage-based perspectives on verb valency 41
tr lm tr
tr lm
kick
In sum, with this model, not only does Langacker provide a plausible account
of how the valency of a verb is tied to the usage of that verb, he also demonstrates
that some extent of construction-related verbal polysemy is actually expected in
a usage-based grammar when the principles whereby verbs and constructions are
taken to combine are examined at close range.
2.4â•… Summary
is assumed that as few lexical entries of a verb as possible should be posited, and
construction-related polysemy (as in projectionist approaches) avoided. I argued
against this assumption on theoretical grounds. I also pointed out that a large body
of evidence emphasizes the need for rich lexical knowledge, which lines up with
the idea that verbs may be stored with more than one lexical entry.
Following a proposal by Langacker (2009), I argued for a usage-based ap-
proach to verb valency, whereby the valency information stored with a verb is
dependent upon prior usage of that verb. In a nutshell, if a verb frequently occurs
in a construction, it will acquire a new lexical entry congruent with that construc-
tion, with the corresponding number of participant roles. Under this account,
more than one construal of the semantic frame may be conventionally associated
with the verb, and the range of lexical entries is not arbitrary but motivated by
usage. In the next chapter, I present evidence for a usage-based account of verb
valency along these lines.
chapter 3
This chapter reports on a case study testing the hypothesis that the valency infor-
mation conventionally associated with a verb is related to the usage of that verb in
different constructions. More precisely, this study investigates whether different
valency patterns of a verb are associated with different degrees of cognitive acces-
sibility determined by the frequency of these valency patterns in usage.
In Section 3.1, I spell out in more detail the usage-based valency hypothesis
and the predictions that it makes. In Section 3.2, I present a reading compre-
hension experiment, in which the relative cognitive accessibility of two differ-
ent valency patterns of three verbs is measured by reading times. In Section 3.3,
I compare the results of this experiment to usage data, by presenting the valency
distribution of these verbs collected from corpora. I conclude that the usage-based
valency hypothesis is largely borne out by my data, although further research needs
to be done to ascertain the determinants of valency with more precision.
The starting point of this case study is the usage-based valency hypothesis, which
I outlined in the previous chapter, and which can be spelled out as follows:
The usage-based valency hypothesis: The cognitive status of a valency pattern of
a verb is related to the frequency of occurrence of that valency pattern with that
verb in usage.
By “valency pattern”, I refer to a given set of participant roles taken from a verb’s
semantic frame; in terms of construal, a valency pattern corresponds to a particular
profiling, or windowing of attention, of a frame. By “cognitive status” of a verb’s va-
lency pattern, I refer to the kind of stored cognitive representation that is involved
in any use of the verb with this valency pattern. As discussed in Chapter 2, valency
patterns can be stored as such in the mental lexicon, or they can be derived com-
positionally when a verb is used in a construction whose semantic specifications
clash with an existing lexical entry. Yet, I do not assume that the cognitive status
46 Argument Structure in Usage-Based Construction Grammar
It has long been observed that in the course of language comprehension, language
users make early parsing decisions that may have to be revised if the following words
conflict with the initial parse, resulting in so-called garden-path effects (cf. Ferreira
& Henderson, 1991). For example, in sentence (1) below, readers are likely to ana-
lyze raced as the main verb, before realizing that it is actually a passive participle in
a reduced relative clause when they encounter the actual main verb (fell). Because
the sentence needs reanalysis, this would translate in increased reading times at fell.
(1) The horse raced past the barn fell.
A large body of research indicates that initial parsing decisions seem to vary from
verb to verb, even when different analyses are a priori equally possible according to
the verb’s subcategorization information. For example, Clifton, Frazier, & Connine
(1984) report that among optionally transitive verbs, some (like teach) are biased
towards a transitive analysis, and others (like perform) are biased towards an in-
transitive analysis, and that these biases correlate with production preferences in
a prior norming study. In a lexical decision task, they found that words signaling a
sentence continuation that is at odds with the verb bias induced a higher processing
cost than words that were compatible with it. For instance, subjects had less dif-
ficulty processing his (signaling the onset of a noun-phrase) after transitive-biased
teach than after intransitive-biased perform in sentence (2a) below, and conversely,
they had more difficulty processing with (signaling that the clause is intransitive)
after teach than after perform in sentence (2b).
(2) a. The aging pianist taught/performed his solo with great dignity.
b. The aging pianist taught/performed with his entire family.
Provided that the omission of the complementizer (that) is allowed, sentences with
such verbs are temporarily ambiguous between an analysis of the post-verbal NP as
the direct object of the verb, or as the subject of a subordinate clause. Using tempo-
rarily ambiguous sentences like (3b) in a cross-modal naming experiment, a self-
paced reading experiment and a reading comprehension experiment in which eye
movements were monitored, Trueswell, Tanenhaus, & Kello (1993) report higher
processing costs for the embedded verb (corresponding to a garden-path effect)
48 Argument Structure in Usage-Based Construction Grammar
when the verb is NP-biased than when it is S-biased. Noting that many of Trueswell
et al.’s stimuli were problematic with regard to the plausibility of a direct object
analysis (for instance, The general pretended the weapon…), Garnsey, Pearlmutter,
Myers, & Lotocky (1997) replicated the experiment by including plausibility as a
controlled factor. Their results line up with Trueswell et al.’s earlier findings and
discard plausibility as a potential confound. Finally, Hare, McRae, & Elman (2003)
provide evidence that verb biases in language comprehension are sense-contingent,
in that senses of the same verb form may have different syntactic preferences. For
example, while find can occur with a direct object or a sentential complement,
when it is used in the sense of ‘locate’ it only occurs with the former type of com-
plement, and when it is used in the sense of ‘realize’, it can occur with either type
but is S-biased. By including contextual information in the stimuli, thus biasing
subjects towards a given sense of the verb, Hare et al. found an effect of verb bias
only when the relevant sense is activated.
Within the context of the present hypothesis, verb biases can be regarded as a
reflection of the fact that different lexical entries of the verb have different degrees
of cognitive accessibility, but how is this related to usage? In all the studies cited
above, verb biases were determined by norming studies, in which subjects had to
produce sentences with a given verb or to complete a sentence prompt containing
that verb. It is not clear to what extent the frequencies gathered in such norm-
ing studies relate to usage. Even though the productions of subjects might be
regarded as usage data in themselves, they are arguably not natural usage, since
they were collected in a controlled setting and were not spontaneous. However,
Lapata, Keller, & Schulte im Walde (2001) do report some correlation between
verb frame frequencies extracted from the BNC using computational learning
techniques and the results of norming studies by Connine, Ferreira, Jones, Clifton,
& Frazier (1984), Trueswell et al. (1993), Garnsey et al. (1997) and Pickering,
Traxler, & Crocker (2000). Yet, they note that the correlation is far from perfect,
and that the predictions of norming studies regarding the bias of particular verbs
do not always match the predictions of the corpus. It is not clear whether such mis-
matches might be due to noise introduced by the automatic learning procedure,
but it is also worth noting that norming studies themselves do not always agree on
these verb biases, as Lapata et al. (2001) observe. Therefore, it would be worthwhile
to compare the performance of norming studies vs. corpus data with respect to
experimental results themselves. Wiechmann (2008) provides such a study by
comparing the results of Hare et al.’s (2003) self-paced reading experiment to usage
data gathered from a subset of the BNC, using the results of distinctive collexeme
analysis (Gries & Stefanowitsch, 2004) as a measure of both form-based and sense-
contingent verb biases. Wiechmann reports a significant correlation of the reading
times collected by Hare et al. (2003) with the corpus-derived sense-contingent
Chapter 3.╇ Empirical evidence for usage-based valency 49
preferences (but not with the form-based ones), and concludes that initial parsing
decisions for these verbs are indeed determined by usage.
In sum, there does seem to be some indication that the psycholinguistic be-
havior of language users in comprehension with respect to argument realization
is related to usage. While these findings are in line with a usage-based account of
grammar in general, we may wonder to what extent they provide evidence for the
usage-based valency hypothesis in particular. In fact, it is not clear whether the
case of NP vs. clausal complementation, which constitutes the bulk of experimental
findings, really qualifies as a change in valency. Indeed, in such cases, neither the
number nor the semantic type of the arguments varies between the two comple-
mentation types, which rather correspond to different realizations of the same
participant role, especially when verb sense (i.e., the semantic frame) is taken into
account. For example, the valency of know, in the sense of ‘be aware of ’, includes
a cognizer and a fact, and the latter may be realized as a noun phrase or a clause,
but the valency pattern itself does not vary between these two realizations. Yet, if
processing differences arise with verb biases related to different realizations of the
same argument, it is likely that they will also arise with biases related to different
sets of arguments. Note that the case of optionally transitive verbs examined by
Clifton et al. (1984) and Pickering et al. (2000) is more in line with the hypothesis,
since it involves one-participant vs. two-participant construals of the same se-
mantic frame. However, the evidence for processing differences in this case is less
abundant, and their frequency basis is disputed by Pickering et al. (2000). Also,
in Lapata et al.’s (2001) study, the transitive vs. intransitive biases measured by
Connine et al. (1984) and Pickering et al. (2000) are generally less well predicted
by the automatically extracted corpus frequencies than the NP/S biases, which
they attribute to the higher rate of tagging and parsing errors. Since the existing
literature does not seem to provide decisive evidence in favor of my hypothesis, I
chose to design and carry out my own experiment, described in the next section.
Since the present study addresses the relation between valency and usage, and since
there is no existing frequency lists of valency patterns in English,1 this information
must be obtained from corpora. This unfortunately puts some constraints on the
material that may be used, and means that, for reasons of time and feasibility, the
study can only be done with a limited number of verbs, which must therefore be
chosen carefully. The three main commerce verbs, buy, pay, and sell, were consid-
ered good candidates for a number of practical as well as theoretical reasons. First,
they are quite frequent, at least frequent enough for me to get a fairly accurate pic-
ture of their valency distribution in usage that is not deformed by modest corpus
size, without having to search through billions of words of running text. Second,
they display rather low polysemy compared to other high- and mid-frequency
verbs, which lowers the need for semantic tagging and makes the annotation pro-
cess easier. As it turns out, these verbs occur most often in their basic sense, and
the other less frequent meanings were fairly easy to identify in the corpus extracts
(cf. Section 3.3.1).
There are certainly other verbs of English that meet these criteria, but buy, pay
and sell present other analytical benefits relating to the question of valency. All three
verbs evoke the commercial transaction frame,2 already mentioned in the previous
1. Some researchers, such as Gahl, Jurafsky, & Roland (2004), do provide corpus-based data
reporting frequency counts of English verb usage. Suitable data can also be found in norming
studies (e.g. Connine et al., 1984), although they are not per se derived from usage. I considered
using such data, but the coding scheme used in these studies is generally limited to purely formal
information (i.e., the syntactic categories of verb complements) and too coarse-grained for my
present purposes. For example, in Connine et al. (1984), all uses of verbs with a direct object and
a prepositional phrase were collapsed into one category “NP PP”, regardless of the preposition
(and hence, of the semantic role of the corresponding argument), and Gahl et al. (2004) did not
at all consider complex transitive uses other than the ditransitive construction (i.e., occurrences
of NP PP were collapsed with plain transitives).
2. This classification is also found in the FrameNet database, albeit with a slight twist. In
FrameNet, buy, pay, and sell are lexical units of the Commerce_buy, Commerce_pay and
Commerce_sell frames respectively. Commerce_buy and Commerce_sell on the one hand, and
Commerce_pay on the other, are perspectivizations of the Commerce_goods-transfer frame
and the Commerce_money-transfer frame respectively, which are themselves subframes (i.e.,
sub-events) of the general Commercial_transaction frame. These refinements in the frame-
to-frame relations and lexical unit assignments are meant to capture semantic subregularities
among commercial transaction events, and to provide a more specific frame-based descrip-
tion of the semantic differences between verbs. At any rate, all three verbs share the general
Commercial_transaction frame at some level of description and ultimately receive their frame
elements and semantics from this frame, as in the much simpler account I offer here.
Chapter 3.╇ Empirical evidence for usage-based valency 51
chapter, which refers to a scenario in which a seller transfers goods to a buyer and re-
ceives money from the latter in return. All four participants (buyer, goods, money
and seller) are necessarily involved in any commercial event, regardless of the verb
that is used to express this event linguistically, and, more importantly, regardless of
the arguments that are actually mentioned in any use of the verb.3 This is what se-
mantically distinguishes, for example, an event of selling from one of giving (which
does not necessarily imply a counter-transfer), or an event of buying from one of
receiving. Therefore, there can be little discussion on purely conceptual grounds as
to whether these four participants can be part of a stored valency pattern for these
verbs, as the argument status of these participants is less controversial than other
cases, such as instruments or locations (cf. Koenig, Mauner, & Bienvenue, 2003).
Finally, these verbs share a distributional property that is interesting for the
purpose of testing the cognitive accessibility of their different valency patterns. All
three are typically used with a direct object argument, referring to the exchanged
goods in the case of buy and sell, and to the exchanged money in the case of pay.
However, other frame elements may be expressed after the direct object, providing
for each verb two possible continuations of the transitive clause. For buy, either the
seller or the money may be expressed, as in (4a) vs. (4b) below. For pay, either the
goods or the seller may be expressed, as in (5a) vs. (5b). For sell, either the buyer
or the money may be expressed, as in (6a) vs. (6b).
(4) a. The tourist bought a phrase book from [the shopkeeper]seller.
Valency pattern: buyer-goods-seller
b. The tourist bought a phrase book for [ten euros]money.
Valency pattern: buyer-goods-money
(5) a. The tourist paid ten euros for [a phrase book]goods.
Valency pattern: buyer-money-goods
b. The tourist paid ten euros to [the shopkeeper]seller.
Valency pattern: buyer-money-seller
(6) a. The shopkeeper sold a phrase book to [the tourist]buyer.
Valency pattern: seller-goods-buyer
b. The shopkeeper sold a phrase book for [ten euros]money.
Valency pattern: seller-goods-money
Each continuation involves a different valency pattern of the verb. This distribu-
tional property can be profitably used to design an incremental reading compre-
hension experiment. Upon reading sentences such as (4a) to (6b) word by word, a
different construal of the frame must be envisaged in order to integrate the content
3. Semantic descriptions of commerce verbs along these lines are provided by Fillmore (1977)
and Croft, Taoka, & Wood (2001).
52 Argument Structure in Usage-Based Construction Grammar
3.2.2â•… Design
––
As argued by Forster (2010), there are two respects in which a maze task is
superior to the classical implementations of the self-paced reading paradigm.
First, self-paced reading makes the assumption that readers synchronize their
key presses with integration processes of single words, but this assumption is
bound to be violated by subjects adopting a “wait-and-see” strategy, whereby
they advance in the sentence at a constant rate and only slow down to integrate
the last couple of words that they have read. Such reading strategies create dif-
ficulties in the analysis of the results, which must be performed on segments of
the sentence rather than on single words in order to reveal systematic differ-
ences in processing time. In a maze task, subjects are forced into a more strictly
incremental mode of processing, because each word must be integrated into
the sentence before moving onto the next word. They must pay attention to the
words they select and cannot make a random selection without running the risk
of being unable to chose between the next two alternatives. Secondly, another
issue with self-paced reading is that the researcher cannot be sure whether the
reader actually processes the sentence accurately and does not tap erratically in
order to get through the task. To avoid this, secondary tasks such as comprehen-
sion questions are usually introduced after each trial, but the reliability of such
controls is often hard to appreciate. A maze task eschews the need for post-trial
controls altogether, since, as already pointed out, subjects cannot reasonably
rely on random key pressing if they are to reach the end of the sentence; as it
were, responses suspected to be unreliable (if any) can actually be weeded out
by removing those trials in which the subject failed to do so. Besides, the activity
required by a maze task, while still unnatural, appears to make more sense to
subjects than mere word by word reading and may well be presented as a game
to the participants, which is likely to promote a higher degree of involvement.
For this reason, a maze task is less prone to elicit parasitic response strategies
than standard self-paced reading.
In sum, the maze task design is tailored to measure processing time at particu-
lar points of a sentence with greater accuracy than a self-paced reading task, which
makes it a prime choice for this experiment, in which the focus is on the integra-
tion time of particular prepositions marking different kinds of participant roles.
3.2.3â•… Stimuli
Three stimuli sentences with the verbs buy, pay and sell were constructed for both
of their valency conditions, each with a different direct object. The direct objects
used with each verb were chosen to refer to widely different kinds of objects (or
different amounts of money in the case of pay): a camera, a painting and a sandwich
for buy, forty euros, ninety euros and ten euros for pay, and his/her bike, his/her
Chapter 3.╇ Empirical evidence for usage-based valency 55
sculpture and his/her watch for sell.4 This variation in the expression of the direct
object was introduced in order to see whether its referential properties had any
effect on the processing of the subsequent argument of the verb. Apart from this,
the stimuli sentences all had a similar structure. The subject was in each case a
proper noun corresponding to a common English first name which was, for each
sentence and each subject, randomly selected from a list balanced for gender. The
direct object consisted of exactly two words, which means that the preposition
occurred as the fifth word in all stimuli. This insured that subjects were under a
similar processing load in all sentences when they encountered the critical word.
After the preposition, a noun phrase was chosen whose referent was a plausible
candidate for the role assigned to it in the commercial transaction frame vis-à-vis
the referent of the direct object (for instance, seventy euros, but not two euros, is
a plausible price for a camera, and an art gallery, but not a pharmacy, is a plausible
seller of paintings). All prices mentioned in the stimuli were given in euros, since I
assumed that it is a currency that all subjects, as students in Germany from differ-
ent English-speaking countries, were familiar with. The eighteen stimuli sentences
(3 verbsâ•›×â•›2 valency conditionsâ•›×â•›3 direct objects) are summarized in Table 3.1.
Table 3.1╇ Stimuli sentences for the maze task experiment, for each verb
and valency condition.
buy for-money Peter bought a camera for seventy euros.
Henry bought a painting for two hundred euros.
Lisa bought a sandwich for three euros.
from-seller Paula bought a camera from the department store.
Jane bought a painting from an art gallery.
Harry bought a sandwich from a takeaway.
pay for-goods Laura paid forty euros for the meat.
Sally paid ninety euros for a necklace.
Terry paid ten euros for a cake.
to-seller Bob paid forty euros to the butcher.
Meg paid ninety euros to the jeweler.
Jill paid ten euros to the baker.
sell for-money Mike sold his bike for seventy euros.
Anna sold her sculpture for one grand.
Mary sold her watch for sixty euros.
to-buyer Bill sold his bike to the neighbor.
Justin sold his sculpture to an old woman.
Susan sold her watch to the landlord.
4. The determiner that was used in the actual sentences depended on the gender of the subject
argument, which, as I am about to indicate, varied across sentences.
56 Argument Structure in Usage-Based Construction Grammar
The stimuli sentences were randomly interspersed with blocks of three filler sen-
tences. It was prevented that two consecutive critical items contained the same verb,
so as to avoid any priming effect that might arise despite the presence of filler items.
The filler sentences were of course constructed with different verbs from those used
in the stimuli, and with different constructions from the critical items they pre-
ceded, also to avoid priming effects. Since the same verb, direct object, or construc-
tion appear more than once in the list of critical items, the filler items were designed
accordingly in order to make the critical items less noticeable, in that they included
sets of sentences with the same verb but a different post-verbal noun phrase and/or
construction. An example of a pair of filler items with the same verb (promise), same
subject and same post-verbal noun phrase but a different construction (ditransitive
vs. that-clause), is given in (7) below. As a matter of fact, this method appeared to
be effective, since subjects rarely mentioned the commerce verbs when they were
asked about the sentences they remembered after the experiment.
(7) a. The contractor promised the workers a pay raise.
b. The contractor promised the workers would finish by Friday.
Each word in all stimuli and filler sentences was matched with a word that did
not make a well-formed continuation with the previous words of the sentence. In
all eighteen stimuli, the preposition was matched with a low-frequency verb, as
determined by the BNC word frequencies reported by Leech, Rayson, & Wilson
(2001).5 While it was undesirable to use the same matching word for the preposi-
tion in all eighteen sentences, lest it be noticed by subjects and have an influence
on their responses, choosing words of the same category and frequency range
insures that discriminating between the correct and the incorrect continuation of
the sentence is roughly equally difficult in all cases, which makes the correspond-
ing reaction times comparable.
Twenty-five native speakers of English (11 male, 14 female) aged between 20 and
35 (24.16 on average), all students at the University of Freiburg, were tested in this
experiment. Most of them (19) came from the United States, four from Canada and
5. These eighteen verbs are: boil, bleed, bless, cater, dive, drip, fade, flee, grasp, greet, peel, plead,
squeeze, steer, trail, vanish, weigh, yawn. To choose them, I consulted the frequency-ranked list
provided by Leech et al. (2001) on the book’s companion website (http://ucrel.lancs.ac.uk/bnc-
freq/flists.html), starting from the bottom of the list and selecting appropriate verbs as I found
them. The frequency of these verbs varies between 4 and 24 per million words (meanâ•›=â•›14.5).
Chapter 3.╇ Empirical evidence for usage-based valency 57
two from England. They were rewarded with a monetary compensation for their
participation, except two who took the experiment for course credit.
After signing a consent form and filling out a questionnaire, the participants
were shown to a computer where the experiment, implemented with the E-Prime
2.0 experiment development environment, took place. Each participant was pre-
sented with a different randomly-generated stimuli list. Following the maze task
design outlined in Section 3.2.2, participants advanced in each sentence incremen-
tally by selecting one of two words shown on the computer screen. They provided
their selection by using a response board with two horizontally placed buttons,
built from a modified computer keyboard, pressing the left button to select the
word displayed on the left side of the screen, and the right button to select the word
displayed on the right. Subjects were instructed to keep their hands on each button
throughout the experiment, so that their response should be as fast as possible. If
their selection was correct, they were presented with the next pair of words; if not,
they were shown an error message and they continued with the next sentence in
the list. The position of the correct word and the incorrect word (left or right) was
randomly determined by the program during the execution. The first word was
always presented on the left side, with a blank (“–”) on the right side, and subjects
could press whichever button in order to proceed to the second word. The whole
task took them about 15 minutes. For each word in each sentence, the program
recorded whether or not the participant made the correct selection, and the time
it took them to make this decision.
3.2.5â•… Results
In this section, I analyze the results of the maze task separately for each verb, using
two dependent measures. The first variable of interest is the reaction time to the
fifth pair of words in the stimuli sentences, which includes the preposition. Trials
in which the subject made an incorrect selection before or at the fifth word were
obviously discarded from the dataset. From the 150 data points collected for each
verb, this resulted in the removal of 8 data points (5.3%) for buy, 4 (2.7%) for pay,
and 8 (5.3%) for sell. Subsequently, I extend the analysis to the sixth word, i.e., the
determiner introducing the noun phrase argument.
Starting with sell, the mean reaction times for each word in the stimuli sen-
tences with that verb are plotted in Figure 3.2. Each valencyâ•›×â•›DO combination is
plotted separately, so that not only the difference between valency conditions but
also the variation of this difference according to the direct object referent (hereafter
DO) can be visualized. The bars on the points of the plot represent the standard
error of each reaction time distribution, and thus provide an indication of how
58 Argument Structure in Usage-Based Construction Grammar
the reaction times vary across subjects compared to the mean value. As indicated
in the legend, DO conditions are contrasted with different dot symbols and differ-
ent shades of gray, and valency conditions are contrasted with different line types
(solid vs. dashed). There appear to be systematic differences in reaction time at the
preposition, with to (indicating a buyer argument) being integrated faster than for
(indicating a money argument) in all three DO conditions.
1200
1000
Mean RT
DO = bike
DO = sculpture
DO = watch
800 to-BUYER
for-MONEY
600
Figure 3.2╇ Mean reaction times to each word in sentences with sell.
To test whether these differences are significant, I submitted the data for sell to
linear regression analysis with mixed effects (cf. Baayen, 2008, Chapter 7; Baayen,
Davidson, & Bates, 2008). As argued by Baayen & Milin (2010), mixed models are
particularly well suited to the analysis of reaction time data, notably because varia-
tion related to random factors, such as subjects (i.e., some subjects may be faster or
slower responders than others, regardless of the stimulus) or items (i.e., some items
may be easier or harder to process regardless of the controlled variables that are
manipulated within them), can be factored in as random effects. For this analysis,
I used the lmer function from the lme4 package in the R environment. Before fit-
ting the data to a mixed model, I normalized the reaction times by using an inverse
transformation. The dependent variable is thus the inverse of the reaction time to
the fifth pair of words of the stimuli sentences (hereafter RTinv). Using inverted
reaction times instead of raw values (in milliseconds) is common in the analysis of
reaction time data. This is motivated by the fact that reaction times do not usually
follow a normal distribution (cf. Luce, 1986) and thus violate the distributional as-
sumptions of many statistical hypothesis-testing techniques. As reported by Ratcliff
(1993), the inverse transformation often normalizes the distribution somewhat and
reduces the effects of outliers, and therefore maintains good power for the analy-
sis. This strategy was also effective in the present study, as shown by Shapiro tests
Chapter 3.╇ Empirical evidence for usage-based valency 59
Table 3.2╇ Effect of Valency in the linear regression analysis for sell.
Model formula: −1000â•›×â•›RTinvâ•›~â•›Valencyâ•›+â•›(1|Subject)â•›+â•›(0â•›+â•›Trial|Subject).
Predictor Estimate MCMCmean Lower Upper p
(Intercept) −1.5695 −1.5626 −1.6887 −1.4287 0.0000
Valency (to-SELLER) −0.2534 −0.2517 −0.3854 −0.1224 0.0002
6. Hence, the formula roughly reads as: “predict the value of RTinv as a function of valency, taking
into account variation across subjects and variation according to the trial number for each subject.”
7. I used the pvals.fnc function from the languageR package.
8. All tables presented in this section use the following abbreviations: MCMCmean: Markov
Chain Monte Carlo mean for the estimated coefficients; Lower, Upper: 95% highest posterior
density intervals; p: p-value based on the t distribution with the number of observations minus
the number of fixed-effects coefficients as degrees of freedom.
60 Argument Structure in Usage-Based Construction Grammar
The very low p-value (< 0.001) indicates that Valency is a highly significant pre-
dictor of reaction time. The model predicts a variation of −0.2534 in −1000â•›×â•›RTinv
when the preposition is to, which corresponds to a decrease of 88.57 msec when
backtransformed.9 In other words, subjects were significantly faster to respond
when the preposition was to than when it was for.
As already noted when commenting on Figure 3.2, the difference in reaction
time between valency conditions seems to hold across all three DO conditions.
However, there appears to be variation in the magnitude of that difference; for
instance, the points for to and for are further apart when the DO is watch than
when it is sculpture. To test whether these differences are significant, I added DO
to the model’s predictors, as well as the interaction between Valency and DO.
The formula for this model, given in the caption of Table 3.3, is identical to the
previous one, except that DO is added as a predictor and checked for interactions
with Valency; this is marked by the ‘×’ sign between the two terms. The results of
the MCMC sampling are summarized in Table 3.3.
Table 3.3╇ Effect of Valencyâ•›×â•›DO in the linear regression analysis for sell.
Model formula: −1000â•›×â•›RTinvâ•›~â•›Valencyâ•›×â•›DOâ•›+â•›(1|Subject)â•›+â•›(0â•›+â•›Trial|Subject).
Predictor Estimate MCMCmean Lower Upper p
(Intercept) −1.5459 −1.5399 −1.7317 −1.3469 0.0000
Valency (to-SELLER) −0.2339 −0.2340 −0.4637 −0.0157 0.0382
DO (watch) −0.1193 −0.1213 −0.1036 −0.3501 0.2978
DO (sculpture) −0.1477 −0.1485 −0.3618 −0.0793 0.1899
Valency (to-SELLER) −0.1035 −0.1055 −0.4136 −0.2324 0.5203
â•›× DO (watch)
Valency (to-SELLER) −0.0350 −0.0365 −0.2913 −0.3404 0.82
â•›× DO (sculpture)
9. I obtained the backtransformed value (in milliseconds) by calculating the difference between
the raw RT value in the for-MONEY valency condition and the raw RT value in the to-SELLER
valency condition. Since, trivially, the inverse of an inverted RT is a raw RT (1 / x −1â•›=â•›x), the for-
mer is obtained by calculating the inverse of the intercept reported in Table 3.2 (−1.5695), which
corresponds to the value of −1000â•›×â•›RTinv at the reference level of valency (for-MONEY), and
multiplying the result by −1000 (−1000â•›×â•›1 / −1.5695â•›=â•›637.15). For the latter, I first calculated the
value of −1000â•›×â•›RTinv in the to-SELLER valency condition by subtracting the effect coefficient
for that level (−0.2534) from the intercept, and applying the aforementioned transformation
to the result (−1000â•›×â•›1 / (−1.5695â•›−â•›0.2534)â•›=â•›548.58). Finally, by subtracting the backtrans-
formed RT values of each level, I obtained the backtransformed RT difference between levels:
637.15â•›−â•›548.58â•›=â•›88.57.
Chapter 3.╇ Empirical evidence for usage-based valency 61
In this new model, the main effect of Valency remains significant (with bike as
the reference level for DO). As can be seen in the last two lines of Table 3.3, neither
of the interactions with the other levels of DO are significant. I obtained the same
results with different values of DO as the reference level. In sum, the facilitating
effect of Valency does not significantly vary with the direct object in this dataset.
I applied a similar analysis to the data for the sentences with the verbs pay and
buy. The mean reaction times to each word in the stimuli sentences with pay are
plotted in Figure 3.3. As with sell, there are differences in the integration time of
the preposition (to, introducing a buyer argument, vs. for, introducing a goods
argument) between valency conditions, but contrary to sell, this difference is only
found in two of the three DO conditions, ninety euros and ten euros.
1000
900
DO = forty euros
800 DO = ninety euros
Mean RT
DO = ten euros
to-SELLER
700 for-GOODS
600
500
paid X euros for/to Det N
Figure 3.3╇ Mean reaction times to each word in sentences with pay.
As with sell, I submitted the data for pay to two regression analyses, one with
Valency as single predictor (Model 1) and one with the interaction between
Valency and DO (Model 2), both with the same random effect structure as pre-
viously. The results of these analyses are presented in Table 3.4. In Model 1, there is
a significant effect of Valency (pâ•›<â•›0.05), whereby the transformed reaction time to
the pair of words containing the preposition varies by 0.1459 when the preposition
is to, corresponding to an increase of 45.18 msec when backtransformed. In other
words, subjects were significantly faster to respond when the preposition was for
than when it was to, if all DO conditions are collapsed. Yet, an examination of the
results of Model 2 unsurprisingly reveals interactions between the two factors.
While there is a significant increase in the to-seller condition at the DO refer-
ence level (here, ten euros), this effect is counterbalanced by a significant decrease
62 Argument Structure in Usage-Based Construction Grammar
when the DO is forty euros: this lines up with what can be observed in Figure 3.3.
The interaction with the level ninety euros is, however, not significant, showing
that the main effect of valency is robust in at least two DO conditions. The reason
for the peculiarity of the DO forty euros, for which virtually no difference at all
in reaction time is found, is unclear. Note, however, that since the interaction is
barely significant, it can arguably be dismissed as a random peculiarity of the data.
Moreover, as will be seen, the interaction actually disappears when the analysis is
extended to the sixth word.
1200
1000 DO = camer
DO = painting
Mean RT
DO = sandwich
from-SELLER
for-MONEY
800
600
Figure 3.4╇ Mean reaction times to each word in sentences with buy.
Chapter 3.╇ Empirical evidence for usage-based valency 63
Finally, the mean reaction times to each word in the stimuli sentences with buy are
plotted in Figure 3.4.10 The picture obtained is much more mixed than for the two
other verbs. For two DOs, the difference in reaction time between valency condi-
tions is very low and points in opposite directions; only for camera is it somewhat
substantial, with faster responses for the preposition for. The difference in mean
RT across DO conditions is virtually nonexistent (from: μâ•›=â•›657.56; for: μâ•›=â•›658.57).
The regression analyses, presented in Table 3.5, lead to similar observations. The
effect of Valency in Model 1 is very weak and not significant. The examination of
interactions reveals a marginally significant effect of Valency at the DO reference
level (i.e., camera), but otherwise no significant interactions. Separate analysis of
each DO condition leads to the same conclusion. In sum, there is no tangible evi-
dence in this dataset for a difference in the integration of the preposition for buy.
To summarize, it was found that the effect of valency on the reaction time to
the pair of words containing the preposition is: (i) highly significant for sell across
all DO conditions, with responses to to faster than to for, (ii) significant for pay in
two of the three DO conditions, (iii) not significant for buy in all DO conditions.
While they do reveal systematic differences in the integration time of the preposi-
tion, these results should nonetheless be taken with a note of caution. So far, I have
10. The peak at Word 7 observed for sandwich in the from-condition is likely to be due to the
unfortunate choice of the noun takeaway, which is apparently only found in British English.
Since most subjects were speakers of North American varieties, they were not familiar with that
word. This is, however, inconsequential for my analysis, since I do not consider reaction times
at the seventh word.
64 Argument Structure in Usage-Based Construction Grammar
focused on the fifth word because it is presumably at this point that subjects must
bring to mind an additional argument of the verb, whose nature depends of the
preposition recognized as the correct continuation of the sentence. However, as
discussed in Section 3.2.1, this assumption must be nuanced due to the fact that the
prepositions used in the stimuli potentially allow other readings than as markers
of arguments of the verbs. These alternate readings can be largely ruled out when
the next word is read, which gives grounds for considering a second dependent
variable that includes the reading time of the sixth word.
I repeated the same linear regression analysis for each verb with the in-
verse of the average of the reading times of the fifth and sixth words in each
trial (again multiplied by −1000) as the dependent variable, viz.â•›−1000â•›×â•›2â•›/
(RT(Word5)â•›+â•›RT(Word6)). As in the first analysis, trials in which the subject made
an incorrect selection before or at the sixth word were discarded from the dataset.
From the 150 data points collected for each verb, this resulted in the removal of
10 (6.7%) for buy, 7 (4.7%) for pay, and 9 (6%) for sell. The results of this analysis
are presented in Table 3.6. As can be seen from the estimates and p-values of the
Valency predictor in each of these tables, the extended analysis yields similar
results to those of the previous analysis. There is a significant (pâ•›<â•›0.01) facilitating
effect for sell when the preposition is to, a highly significant (pâ•›<â•›0.0001) hindering
effect for pay when the preposition is to, and still no significant effect for buy. There
was no significant interaction between Valency and DO (not reported here) for
any of the three verbs, whichever reference level was chosen for DO. In particular,
the interactions that were previously found with pay disappear in this new analysis.
Table 3.6╇ Effects of Valency for sell, pay, and buy with the inverse mean of the reaction
time to the fifth and sixth words as dependent variable.
sell
Predictor Estimate MCMCmean Lower Upper p
(Intercept) −0.6618 −0.6602 −0.7219 −0.5989 0.0000
Valency: to-BUYER −0.0881 −0.0890 −0.1433 −0.0357 0.0014
pay
Predictor Estimate MCMCmean Lower Upper p
(Intercept) −0.8863 −0.8817 −0.9468 −0.8222 â•⁄↜0.0000
Valency: to-SELLER −0.1480 −0.1470 −0.0871 −0.2048 <0.0001
buy
Predictor Estimate MCMCmean Lower Upper p
(Intercept) −0.7062 −0.7043 −0.7665 −0.6451 0.0000
Valency: from-SELLER −0.0289 −0.0256 −0.0766 −0.03 0.25
Chapter 3.╇ Empirical evidence for usage-based valency 65
In the next section, I compare these results with frequency data collected from
corpora. I will then turn to the interpretation of these findings with respect to the
usage-based valency hypothesis.
The maze task experiment reported on in the last section revealed the following
differences in processing between pairs of valency patterns of the verbs sell, pay
and buy, as measured by reading times:
In this part of the case study, my purpose is to evaluate the frequency with which
speakers of English witness the verbs buy, pay and sell with different kinds of
complements. For this purpose, a corpus that is representative of the linguis-
tic experience of English speakers is needed. As many scholars have noted, the
issue of how texts of different genres should be balanced in order to constitute a
representative sample of the linguistic experience of speakers is a difficult one,
which is probably impossible to resolve in absolute terms (Biber, 1993; Habert,
2001). There is, however, one kind of discourse that arguably constitutes most of
what speakers do with their language: face-to-face spoken interactions. For that
reason, instead of using a corpus containing an arbitrarily balanced collection of
66 Argument Structure in Usage-Based Construction Grammar
different kinds of texts, I chose to derive usage data from a corpus of conversa-
tions. I extracted all instances of the verbs buy, pay and sell from the 4-million
words conversation part of the XML edition of the British National Corpus. This
part of the corpus comprises 3,754 conversations between speakers of British
English from different parts of Great Britain, recorded in various settings (mostly,
but not exclusively, in private homes).
While the composition and the more than reasonable size of this corpus makes
it a good candidate for this study, it may be objected that it does not really reflect
the linguistic usage experienced by the participants to my experiment, since most
of them (19 out of 25) were speakers of American English, which I did not antici-
pate when I started this research. While there is a priori no reason to expect that
speakers of different varieties of English would differ in the frequency with which
they mention the arguments of these verbs (especially as far as the valency patterns
examined in my experiment are concerned), this is a possibility that cannot be
ignored. Therefore, comparable American English data must also be considered.
As there is to this day no corpus of conversational American English that is com-
parable in size and content to the conversation part of the BNC,11 I decided to use
a combination of various freely accessible corpora: the Santa Barbara Corpus of
Spoken American English12 (SBCSAE), 13 texts from the Saarbrücken Corpus of
Spoken English13 (SCSE) recorded in the United States, the CallFriend corpus of
telephone conversations, and the Charlotte Narrative and Conversation Collection.
The first three were downloaded from the TalkBank website,14 the last one was
taken from the open version of the American National Corpus.15 The resulting
compilation totals about 600,000 words in 194 texts. While it is of a relatively mod-
est size compared to the British corpus (about 15%), it is sufficient for the purpose
of checking whether there are any substantial differences in the relative importance
of the valency patterns of each verb between British and American English.
11. To the exception of the 3-million words Switchboard corpus of telephone conversations,
but I discarded this resource because it does not really contain spontaneous interactions, since
the topic of conversation was imposed in each case. As a matter of fact, preliminary searches in
this corpus revealed that the imposed topic does substantially influence how buy, pay and sell
are used by speakers. For example, with the topic ‘cars’, speakers were more likely to include the
money argument, especially when they meant to tell what a good bargain they made.
12. http://www.linguistics.ucsb.edu/research/sbcorpus.html
13. http://www.uni-saarland.de/fak4/norrick/scose.htm
14. http://talkbank.org/
15. http://www.americannationalcorpus.org
Chapter 3.╇ Empirical evidence for usage-based valency 67
From the British corpus, I obtained 3,634 tokens for buy, 2,926 tokens for
pay, and 1,030 tokens for sell.16 I selected a random sample of 50% of the in-
stances of each verb. From the American corpus, I obtained 274 tokens for buy,
223 tokens for pay, and 107 tokens for sell, which were included in the analysis.
In addition to the utterance in which each token was found, I also included parts
of its discourse context by extracting the four last and the three next utterances.
When further context was needed, I consulted the original text by querying the
online version of the BNC hosted at Lancaster University17 or the original XML
files for the American corpus. For each token, I identified the arguments of the
verb (buyer, goods, money, seller) in the surrounding text18 and assigned the
relevant participant role to each of them. I also included beneficiaries (which were
particularly frequent with the verb buy), labeled “recipient” in this study (since
they are intended recipients as per Goldberg, 1995), whether they were realized
as a for-PP or as an indirect object. To qualify as an argument of the verb in my
coding scheme, a constituent also had to be in a syntactic dependency relation to
the verb, which covers two cases:
– The complement is in the same clause as the verb, whether in its “canonical”
position (e.g., after the verb for an object argument), or displaced to a non-
canonical position because of a particular clause construction, such as a rela-
tive clause or an information structure construction (e.g., preposing).
– The verb and the complement are in different clauses of the same sentence, but
the identity of the “missing” complement is syntactically determined. A com-
mon case falling in this category is that of catenative constructions, for which
there is no indeterminacy as to the identity of the understood subject or object
of the verb in the embedded non-finite clause. For example, in the sentence
he wanted to buy a truck, the pronoun he would be tagged as the subject of the
embedded verb buy. To-infinitive purpose adjuncts also meet these criteria.
I ignored interruptions by another speaker, as well as overlaps, and considered
interrupted utterances as a single sentence. Also, during the annotation phase,
I discarded tokens that do not evoke the commercial transaction frame or were
not verbs, as well as incomplete utterances (such as false starts or hesitations),
utterances containing material that could not be transcribed by the editors of the
16. I used the CQP corpus query tool (http://cwb.sourceforge.net/) to index and access both
corpora.
17. http://bncweb.lancs.ac.uk/
18. I used the GATE Developer text engineering environment (http://gate.ac.uk) to insert an-
notations in running text.
68 Argument Structure in Usage-Based Construction Grammar
Table 3.7╇ Distribution of discarded and retained tokens from each corpus.
British English
buy pay sell
OK 1544 (85%) 1178 (80.4%) 439 (85.3%)
Not a verb â•⁄â•⁄â•⁄0 (0%) â•⁄â•⁄22 (1.5%) â•⁄â•⁄0 (0%)
Other frame â•⁄â•⁄28 (1.5%) â•⁄â•⁄67 (4.6%) â•⁄ 14 (2.7%)
Incomplete â•⁄â•⁄83 (4.5%) â•⁄â•⁄45 (3%) â•⁄ 19 (3.7%)
Unclear â•⁄ 162 (9%) â•⁄ 153 (10.5%) â•⁄ 43 (8.3%)
Total 1817 1465 515
American English
buy pay â•⁄ sell
OK 259 (94.5%) 171 (76.7%) â•⁄ 99 (92.5%)
Not a verb â•⁄â•⁄0 (0%) â•⁄ 10 (4.5%) â•⁄â•⁄2 (1.9%)
Other frame â•⁄â•⁄1 (0.4%) â•⁄ 27 (12.1%) â•⁄â•⁄2 (1.9%)
Incomplete â•⁄ 13 (4.7%) â•⁄â•⁄9 (4%) â•⁄â•⁄3 (2.8%)
Unclear â•⁄â•⁄1 (0.4%) â•⁄â•⁄6 (2.7%) â•⁄â•⁄1 (0.9%)
Total 274 223 107
corpus (which might have been a complement of the verb), and tokens in which
the assignment of participant roles of the verb to surrounding constituents was
unclear, even when further contextual information was consulted. The distribution
of discarded and retained tokens from each corpus is summarized in Table 3.7.
This section reports the valency distributions obtained from the corpus for each
verb. The presented figures correspond to the frequency with which each verb oc-
curs with a given set of overtly realized participant roles. I excluded passive uses
from the counting, on the grounds that passives are structurally different from
the active sentences used in the experiment. Since passive uses only represent a
small number of instances, discarding them does not change the figures drasti-
cally anyway. By the same token, only utterances with an overt subject were kept.
This was because there are several types of subjectless clauses in English, such as
imperatives and declaratives in which the unexpressed subject is understood to be
the speaker, and not all of these constructions necessarily have to do with different
construals of the semantic frame.
Chapter 3.╇ Empirical evidence for usage-based valency 69
Sell
The valency distribution of sell is presented in Table 3.8. Sentences exemplifying
the most common valency patterns are listed in Table 3.9 as an illustration. In each
case, I took one sentence from the British corpus and, if available, one sentence
from the American corpus. The source of the sentence (corpus, text and position
in the text) is given in brackets.
As can be seen from Table 3.8, sell is primarily transitive, as its most common va-
lency contains two participant roles, seller, realized as the subject, and goods,
realized as the direct object. There does not seem to be substantial differences in
distribution between British English and American English, as the most promi-
nent valency patterns are the same in both varieties. Indeed, a high and signifi-
cant correlation is found between the two distributions (Spearman’s ρâ•›=â•›0.7578,
pâ•›=â•›0.01). Concerning the valency patterns examined in the experiment, marked
in boldface in Table 3.8, seller-goods-buyer is significantly more frequent than
seller-goods-money, in both varieties (BrE: χ2(1)â•›=â•›24.6533, pâ•›<â•›0.0001; AmE:
χ2(1)â•›=â•›15.2105, pâ•›=â•›0.0001). This difference in frequency correlates with the higher
cognitive accessibility of seller-goods-buyer in sentence comprehension.
Pay
The valency distribution of pay, presented in Table 3.10, is less skewed than that of
sell. Two valency patterns are roughly equally prominent, buyer-money, realized as
a transitive clause, and buyer-goods, typically realized as an intransitive clause with
a for-PP. The ranking of valency patterns differs slightly between varieties, but the
overall importance of each pattern is relatively similar. A very high and significant
70 Argument Structure in Usage-Based Construction Grammar
Buy
The valency distribution of buy is presented in Table 3.12. Examples of typical
sentences for each valency pattern are listed in Table 3.13. Like sell, buy is primarily
transitive, with the buyer argument realized as subject and the goods argument
realized as direct object. The verb is also very commonly used with a recipient
argument, referring to the intended recipient of the bought goods. There are no
notable differences between varieties, as the top valency patterns are ranked in a
similar order in both corpora. A high and significant correlation is found between
the two distributions (Spearman’s ρâ•›=â•›0.86; pâ•›=â•›0.0002).
Chapter 3.╇ Empirical evidence for usage-based valency 71
make the right predictions in this case. However, the frequency difference is mark-
edly weaker here than between the relevant valency patterns of pay and sell: for buy,
the more frequent of the two tested patterns is only 1.56 times more common than
the less frequent one, whereas this ratio is 3.69 for sell and 2.58 for pay. Besides,
the p-value of the chi-square test for the frequency difference is barely below the
significance threshold of 0.05 in both varieties. It might thus be the case that both
valency patterns have roughly the same cognitive accessibility, which means that
they might not differ enough to reveal a measurable difference in processing time.
Under this understanding, the results of the experiment are perhaps not at odds
with those of the corpus study.
Moreover, it is worth pointing out that the frequencies on which this analysis
is based are absolute frequencies, and that it should be considered whether the
relative importance of these frequencies vis-à-vis the overall frequency of the verb
could also be relevant. In that connection, the distribution of buy is markedly more
skewed than that of pay or sell: while the frequencies of the two valency patterns
under study do seem to differ, they both are very small compared to the overall
frequency of buy (cf. the “%” columns), while the relevant valency patterns of pay
and sell are comparatively relatively common among the uses of these verbs. This
difference in relative frequencies might indicate that the two three-participant
uses of buy are both less familiar to speakers than those of pay and sell, and might
therefore always be derived compositionally.
Certainly, further research is needed in order to confirm this interpretation. In
the meantime, the usage-based valency hypothesis does seem to receive empirical
support, which the case of buy does not clearly compromise. I recapitulate this
evidence in the concluding section of this chapter, while also discussing possible
refinements and offering prospects for future research. Before turning to this, I
would like to bring up a final point concerning the frequency determinants of
cognitive accessibility. The frequency counts presented in this section are based on
which arguments of the verb are realized, but do not take the linguistic realization
of these arguments into account; for example, the buyer argument of sell can be
realized either as an indirect object, as in He sold her a car, or as a to-PP, as in my
stimuli. This counting strategy is in line with the model outlined in the previous
chapter, where repeated occurrence of a verb in constructions profiling a given
valency pattern leads to the conventionalization of this valency pattern with that
verb. However, we may still wonder what kind of frequency best correlates with
cognitive accessibility: the frequency of occurrence of the verb in a particular con-
struction, or the frequency of sets of participant roles, regardless of their realiza-
tion. To address this question, I counted how many instances of the tested valency
patterns of each verb match the specific syntactic patterns used in my experiment.
The resulting frequencies are presented in Table 3.14.
74 Argument Structure in Usage-Based Construction Grammar
Table 3.14╇ Frequency of each verb in the syntactic patterns used in the experiment.
Syntactic pattern F (BrE) F (AmE)
seller sell goods to buyer 37 12
seller sell goods for money 12 1
buyer pay money for goods 85 16
buyer pay money to seller 5 0
buyer buy goods from seller 33 14
buyer buy goods for money 13 3
As can be seen from Table 3.14, the same frequency relations hold whether formal
realization is taken into consideration or not. Sell occurs more frequently in the
pattern ‘seller sell goods to buyer’ than in ‘seller sell goods for money’ and pay
occurs more frequently in ‘buyer pay money for goods’ than in ‘buyer pay money
to seller’. These differences line up with the contrasts in cognitive accessibility
measured in the experiment. As with the form-independent frequencies, the dif-
ference between occurrences of buy in ‘buyer buy goods from seller’ and ‘buyer
buy goods for money’ is weak, although it approaches that of sell. In sum, these
data do not provide a clear answer to the question of whether formal realization
plays a role in the usage basis of valency. I return to this in my concluding words.
3.4â•… Conclusion
In this chapter, I tested the usage-based valency hypothesis, whereby the cognitive
status of the valency properties of a verb is related to the usage of that verb. The
hypothesis is based on the premise that the various valency patterns of a verb can be
stored as such in the mental lexicon or can be derived compositionally by combining
another valency pattern with particular constructions, with intermediate degrees of
emancipation. I suggested that as a valency pattern of a verb becomes more eman-
cipated, it also becomes more cognitively accessible for subsequent uses. I proposed
integration time in language comprehension as a measure of cognitive accessibility,
and I presented a maze task experiment designed to evaluate the relative cognitive
accessibility of two different valency patterns of three verbs evoking the commercial
event frame, containing the participant roles buyer, goods, money and seller:
buy, pay, and sell. By measuring differences in the incremental integration time of
these valency patterns, I found that (i) for sell, the valency pattern seller-goods-
buyer, as exemplified by Bill sold his bike to the neighbor, is more cognitively ac-
cessible than the valency pattern seller-goods-money, as exemplified by Bill sold
his bike for seventy euros, (ii) for pay, the valency pattern buyer-money-goods, as
exemplified by Laura paid forty euros for the meat, is more cognitively accessible
Chapter 3.╇ Empirical evidence for usage-based valency 75
“the more frequently children hear a verb used in a particular construction (the
more firmly its usage is entrenched), the less likely they will be to extend that
verb to any novel construction with which they have not heard it used”. While
Tomasello’s proposal concerns argument structure productivity, i.e., the tendency
for learners to combine verbs with constructions creatively, it can naturally extend
to the question of the cognitive status of attested yet infrequent uses of a verb.
Secondly, I also begged the question of whether the usage determinant of
valency could in fact be the frequency of the verb in a particular syntactic pattern,
and not just the frequency of the verb with a set of participant roles, irrespective of
their syntactic realization. This question directly bears on the kind of information
that is assumed to be stored in the lexical entries of verbs, and in particular whether
it should include information about the formal realization of arguments, and not
just an array of participant roles. The literature on NP/S ambiguities, discussed in
Section 3.1.2, does show that the frequency of particular syntactic realizations of
the same argument of the verb has an influence on the processing of these uses of
the verb. These findings show that verbs must, at some level of storage, make refer-
ence to formal realization, which could be modeled by verb-specific constructions
(Croft, 2003, cf. also Boas, 2003; Iwata, 2008) in a usage-based grammar. This idea
echoes Herbst’s (2011) proposal that grammar must contain a great deal of valency
constructions, i.e., verb-specific instantiations of argument structure construc-
tions, as mentioned in Section 2.2.2. The question of how such constructions would
interact with a level of form-independent lexical entries, if such a level needs to be
maintained, remains to be examined.
In conclusion, this case study shows that the usage-based valency hypothesis
appears to be valid, although I must acknowledge that much research is yet to be
done in order to define the relation between the usage of verbs and their cognitive
representation with more precision. Of prime relevance would be to extend the
analysis to more verbs and to different languages, and to control for other factors.
Other experimental methods to assess the cognitive accessibility of particular uses
of a verb should also be tried in order to provide converging evidence. At any
rate, this study shows the benefits of studying valency issues from a usage-based
perspective.
part ii
Constructions
chapter 4
Part I of this book essentially argued that the structure of the lexical entries of verbs
depends on the occurrence of these verbs in particular constructions. Under this
view, constructional meaning, which selects and structures the conceptual content
of the semantic frames evoked by verbs (including the number of expressed argu-
ments), is instrumental in shaping the verbal lexicon. But where does it come from?
This chapter discusses the idea that the meaning of a construction is extracted from
the meaning of verbs prominently occurring in that construction.
In Section 4.1, I outline the hypothesis of the lexical origin of constructional
meaning and present existing empirical evidence for it. In Section 4.2, I point to
the limits of the hypothesis with respect to constructions with a highly abstract
meaning, such as the conative construction, for which the relation between con-
structional meaning and verbal distribution is much more tenuous than for previ-
ously studied constructions.
1. It should be noted that while this definition applies to argument structure constructions, it
might not equally do to other kinds of constructions (see Fillmore, Lee-Goldman, & Rhodes,
2012 for a typology of constructions). As one of the anonymous reviewers points out, the con-
structions of a language may vary greatly in the amount of meaning that they convey, and
80 Argument Structure in Usage-Based Construction Grammar
Evidence for the connection between the meaning of a construction and its verbal
usage comes from the finding that the verbal distribution of many constructions
is biased towards one particular verb whose meaning closely corresponds to the
meaning of the construction. Searching for the developmental basis of construc-
tional meaning, Goldberg, Casenhiser, & Sethuraman (2004) analyze the distri-
bution of three formally defined syntactic patterns in the speech of 15 mothers
according to some accounts, some constructions do not even convey meaning at all (cf. Fillmore,
1999, but see Goldberg, 2006 for an alternative analysis). Since I am dealing with argument
structure in particular, I will use the term “construction” as a shorthand for “argument structure
construction” in the remainder of this chapter.
Chapter 4.╇ The usage basis of constructional meaning 81
addressed to children in the Bates, Bretherton, & Snyder (1988) corpus: (i) the
VL pattern, consisting of a verb followed by a locative complement,2 (ii) the VOL
pattern, consisting of a verb followed by an object NP and a locative complement,
and (iii) the VOO pattern, consisting of a verb followed by two NPs (the so-called
double-object pattern). They find that the verbs most frequently occurring in these
syntactic patterns are respectively go, put and give, as reported in Table 4.1.
As it turns out, the meaning of these verbs closely corresponds to the central mean-
ing of an argument structure construction with the relevant pattern as syntactic
pole: respectively the intransitive-motion construction, conveying motion along
some path, the caused-motion construction conveying caused change of location,
and the ditransitive construction, conveying caused change of possession.
Goldberg et al. (2004) argue that the distributional biases of constructions
play a crucial role in the acquisition of argument structure. The corpus study of
child-directed speech reveals that the most frequent verb in each construction is
semantically general and part of the basic vocabulary; unsurprisingly, they are
among the most frequent words of the language. Many language acquisition stud-
ies document that such “general-purpose verbs” as go, put and give are among
the first verbs acquired in a number of languages (cf. Clark, 1978). Moreover,
these verbs encode scenes that are directly relevant to human experience, they
are therefore more likely to be grasped by children at an early age and to facilitate
their acquisition of argument structure semantics. These facts lead Goldberg et al.
(2004) to hypothesize that “it is the high frequency of particular verbs in particular
constructions that allows children to note a correlation between the meaning of a
particular verb and the pattern itself, giving rise to an association between meaning
and form” (p.â•›299). Under this account, the meaning of argument structure con-
structions is derived from the meaning of verbs in the construction’s distribution;
the high frequency of general-purpose verbs gives them a preponderant role in
this process of abstraction, resulting in the meaning of these verbs being directly
associated with the construction.
2. I.e., “a preposition phrase indicating location, a particle indicating location (e.g., down, in),
a locative (there, here), or some combination” (Goldberg et al., 2004, p.â•›294).
82 Argument Structure in Usage-Based Construction Grammar
It should be noted that such biases are not an artefact of the corpus of child-
directed speech, even though this result has been initially obtained from such data.
The same distributional properties can also be observed in adult speech, and are
likely to occur to various extents in all kinds of genres. Table 4.2 presents the most
frequent verb found in the distribution of five argument structure constructions
in the spoken part of the ICE-GB corpus, reported by Perek & Lemmens (2010).
Table 4.2╇ Distributional biases of five constructions in adult speech found in the spoken
part of the ICE GB; source: Perek & Lemmens (2010, §9).
Construction Meaning Most frequent verb
Intransitive-motion (Subj V Loc) X go Y go (32%)
Caused-motion (Subj V NP Loc) X cause [ Y go Z ] put (24%)
Ditransitive (Subj V NP NP) X cause [ Y have Z ] give (50%)
Intransitive-resultative (Subj V Cpl) X become Y become (38%)
Resultative (Subj V NP Cpl) X cause [ Y become Z ] make (40%)
Further evidence that the meaning of a construction is derived from its verbal
distribution is provided by a series of experiments conducted by Goldberg et al.
(2004). In their study, 81 undergraduate students had to learn a novel argument
structure construction on the basis of a minimal input. This novel construction
associates a scene of appearance (some entity appears at some location) with a
syntactic pattern unused in English, namely the SOV pattern (with the theme as
subject and the location as object). In order to acquire the novel construction,
subjects were presented with 16 video clips (showing scenes of appearance) paired
with a linguistic description using the novel construction, instantiating one of
five nonce verbs which all contained a distinctive word-final -o morpheme. Each
verb was associated with a stable meaning, viz. appearance in some manner (see
Goldberg & Casenhiser, 2006, p.â•›195 for the detailed data). Subjects were trained
in either of two conditions. In the balanced condition, all five verbs were used with
relatively low frequency (i.e., 4-4-4-2-2), and no particular verb was used more
frequently than the others. In the high-token-frequency condition, one of the verbs
was used much more frequently than the others; in order words, the construction
was distributionally biased towards this verb in the input given to subjects (i.e., 8-2-
2-2-2). To provide a ‘realistic’ input resembling child-directed speech (in which
the most frequent verb in a construction is semantically general), this verb was
intended to encode a maximally general scene of appearance, without designating
a particular manner of appearance. A third group of subjects who did not receive
this training was taken as control.
After the training phase, subjects were tested to determine whether they ac-
quired the semantics of the construction. They were presented with two scenes, a
scene of appearance and another involving a related action, and were asked which
scene best matched the linguistic description that they were hearing. As expected,
subjects with no prior training did no better than chance at this test. Subjects
in the balanced condition performed significantly better than chance, showing
that they indeed acquired the meaning of the novel construction, but subjects in
the high-token-frequency condition performed even better. This shows that the
skewed input which was presented to them had a significant facilitatory effect on
their acquisition of the construction. This effect can be taken as evidence that the
meaning of a construction is indeed derived from the meaning of verbs frequently
occurring in that construction. If it was not the case, increasing the frequency
of the verb whose meaning most closely correspond to that of the construction
would not be expected to have such a facilitatory effect on the acquisition of that
construction.
84 Argument Structure in Usage-Based Construction Grammar
Table 4.3╇ Contingency table for collexeme analysis. Legend: F(V ∧ C)â•›=â•›frequency
of the verb in the construction, F(¬V ∧ C)â•›=â•›frequency of construction with other verbs,
F(V ∧ ¬C)â•›=â•›frequency of the verb in other constructions, F(¬V ∧ ¬C)â•›=â•›frequency
of the other constructions with other verbs.
Verb V Other verbs
Construction C F(V ∧ C ) F(¬V ∧ C) F(C )
Other constructions F(V ∧ ¬C ) F(¬V ∧ ¬C) F(¬C )
F(V) F(¬V)
Chapter 4.╇ The usage basis of constructional meaning 85
3. Despite the wide range of available distributional statistics, Stefanowitsch & Gries (2003)
argue that Fisher’s exact test is a perfect choice for collostructional analysis, since it “neither
makes any distributional assumptions, nor does it require any particular sample size” (p.â•›218).
4. Stefanowitsch & Gries’ (2003) original paper presents the collostruction strength scores in
their ‘raw’ p-value format. I converted the original values found in the paper to the more read-
able format suggested by Gries & Stefanowitsch (2004), as discussed above.
86 Argument Structure in Usage-Based Construction Grammar
the strongest collexeme of the construction, which is expected given the principle
of semantic compatibility: among the many ways in which a verb can be compat-
ible with a construction, give and the ditransitive exemplify the optimal case where
there is semantic identity. In other words, since the verb give is maximally compat-
ible with the ditransitive construction, it comes as no surprise that it is its strongest
collexeme. Among the thirty top collexemes of the construction, six instantiate the
central meaning of the construction (i.e., actual transfer: give, send, offer, lend, pay,
hand), and four more arguably instantiate metaphorical extensions thereof (e.g.,
transfer of a message, of knowledge, etc.: tell, show, teach, award). It is true, as the
authors argue, that the basic ‘transfer’ sense of the ditransitive is not overwhelmingly
dominant in the collexemes of the construction in terms of verb types, in that there
are relatively few significant collexemes instantiating the central sense in the whole
list. Rather, the high diversity of verbs provides, according to Stefanowitsch and
Gries (2003), evidence for the polysemy analysis of the construction put forward by
Goldberg (1995). That being said, these few verb types are clearly clustered towards
the top of the list: at least four of them (eight including the metaphorical uses) are
among the ten top collexemes. Hence, the collexeme analysis of the ditransitive
construction shows that the verbs most attracted by a construction are those whose
meaning most closely corresponds to the core meaning of that construction. The
authors conclude that “[it] seems that strong collexemes of a construction provide
a good indicator of its meaning” (Stefanowitsch & Gries, 2003, p.â•›227).
Table 4.4╇ The thirty most attracted collexemes of the ditransitive construction
in the ICE-GB corpus; source: Stefanowitsch & Gries (2003, p.â•›229). The token frequency
of each verb in the construction is given in brackets. CollStrâ•›=â•›collostruction strength.
Rank Collexeme CollStr Rank Collexeme CollStr
1 give (461) ∞ 16 allocate (4) 5.54
2 tell (128) 126.80 17 wish (9) 5.51
3 send (64) 67.14 18 accord (3) 5.09
4 offer (43) 48.48 19 pay (13) 4.63
5 show (49) 32.65 20 hand (5) 4.52
6 cost (20) 21.95 21 guarantee (4) 4.33
7 teach (15) 15.36 22 buy (9) 4.2
8 award (7) 10.86 23 assign (3) 3.58
9 allow (18) 9.95 24 charge (4) 3.52
10 lend (7) 8.55 25 cause (8) 3.25
11 deny (8) 8.35 26 ask (12) 3.2
12 owe (6) 7.57 27 afford (4) 2.97
13 promise (7) 7.49 28 cook (3) 2.48
14 earn (7) 6.67 29 spare (2) 2.46
15 grant (5) 5.88 30 drop (3) 1.67
Chapter 4.╇ The usage basis of constructional meaning 87
For another example, a similar pattern of results can be found in Gries, Hampe, &
Schönefeld’s (2005) collexeme analysis of the as-predicative construction (Subj V
NP as-PP), e.g., She described herself as an honest person. In a preliminary analysis
of this construction, Gries et al. suggest that, at the most general level, this con-
struction “expresses the subject’s epistemic stance towards the (atemporal) relation
between the entities referred to by the direct object, and the entities, properties or
states-of-affairs referred to by the as-complement” (p.â•›640). They show that this
semantic description is likely to be the central meaning of the construction, since
it is through various extensions or elaborations of this meaning that other classes
of verbs may be used in the construction. For example, verbs of speech acts, such
as describe, define and denounce, instantiate this central meaning when they occur
in the as-predicative construction, and at the same time elaborate it by making
more explicit how the subject referent expresses its epistemic stance (i.e., within
some kind of speech act, cf. Gries et al., 2005, p.â•›653).
The results of Gries et al.’s (2005) collexeme analysis of the as-predicative con-
struction are reproduced in Table 4.5. They find that regard is the strongest col-
lexeme of the construction. It is also one of the most basic member of a cluster of
mental verbs, many of which are ranked high in the list of collexemes: e.g., regard,
know, recognize and consider. Verbs of perception such as see, view, and perceive,
in their cognition reading, may also be included in this class. In sum, as in the
case of the ditransitive construction, the verbs that most prominently occur in the
as-predicative construction (and especially the top collexeme regard) closely cor-
respond to the hypothesized central meaning of the construction. As Gries et al.
(2005) put it, “the attested usage of regard also implies that, once this verbal item
is known, it will most strongly be associated with the as-predicative, representing,
as it were, the compressed version of the construction’s semantics” (p.â•›652).
This study also provides evidence that collostruction strength is more appro-
priate than raw frequencies for studying the usage of constructions. The authors
observe that the top collexeme regard is only ranked third in terms of raw fre-
quency, while the third strongest collexeme see is actually the most frequent verb.
Gries et al. (2005) argue that this discrepancy between a token-frequency-based
ranking and a collostruction-strength-based ranking shows that, at least in this
case, collostruction strength provides better results than potential misleading raw
frequency. Because see is highly polysemous and is more frequently used as a verb
of perception than as a verb expressing epistemic stance, “the clue it provides for
the description of the constructional meaning of the as-predicative is only vague”
(p.â•›650). It might even be argued that the ‘epistemic stance’ reading of see is con-
tributed by the as-predicative construction and not necessarily by the verb itself.
Gries et al. conclude that “it appears counter-intuitive to base one’s analysis of the
as-predicative’s semantics more on see than on regard” (p.â•›652), thus invalidating
88 Argument Structure in Usage-Based Construction Grammar
Table 4.5╇ The thirty most attracted collexemes of the as-predicative construction
in the ICE-GB corpus; source: Gries et al. (2005, p.â•›649).
Rank Collexeme CollStr Rank Collexeme CollStr
1 regard (80) 166.48 16 class (3) 5.92
2 describe (88) 134.87 17 denounce (3) 5.38
3 see (111) 78.79 18 dismiss (4) 5.16
4 know (79) 42.80 19 consider (9) 5.08
5 treat (21) 28.22 20 accept (7) 4.47
6 define (18) 23.84 21 name (4) 4.28
7 use (42) 21.43 22 portray (3) 3.96
8 view (12) 17.86 23 advert to (4) 3.84
9 map (8) 12.80 24 diagnose (2) 3.44
10 recognize (12) 12.16 25 think of (6) 3.21
11 categorize (6) 11.53 26 depict (2) 3.17
12 perceive (6) 8.30 27 cite (2) 3.06
13 hail (3) 6.32 28 rate (2) 3.06
14 appoint (5) 6.07 29 train (3) 2.98
15 interpret (5) 6.07 30 cast (3) 2.95
The research reviewed in this section provides substantial evidence for a close
connection between the meaning of a construction and its verbal distribution:
1. Many constructions are distributionally biased towards one verb whose mean-
ing is very similar (if not identical) to that of the construction (cf. Goldberg et
al., 2004; Perek & Lemmens, 2010).
2. There is experimental evidence that distributional information plays a role
in the acquisition of constructions, in that a construction is learned more
easily both by adults and children if this construction is distributionally bi-
ased towards a verb with the most general meaning (cf. Goldberg et al., 2004;
Casenhiser & Goldberg, 2005).
3. In line with the research on distributional biases, corpus-based studies using
the method of collostructional analysis report that, for many constructions,
the verbs most closely matching the constructional meaning are among the
strongest collexemes (cf. inter alia Gries & Stefanowitsch, 2003; Gries et al.,
2005). In particular, the top collexeme is typically the most semantically basic
of these verbs and thus the most similar to the constructional meaning.
This body of research strongly suggests that the acquisition of constructional
meaning may well be distributionally based, in that speakers infer the meaning
ascribed to a syntactic pattern from the meaning of verbs frequently occurring in
that pattern. In a usage-based perspective, constructions can thus be described
as statistical correlations of a syntactic form with a lexical meaning, stored as a
form-meaning pair. It is this lexically-derived meaning that determines the pro-
ductivity of the construction, i.e., with which verbs the construction might be used
creatively. Verbs that are semantically similar to the constructional meaning are
allowed in the construction. Verbs that are semantically unrelated may be used
in the construction if there is a way to integrate the meaning of the verb into the
constructional meaning. In that case, the constructional meaning is contributed
by the construction.
90 Argument Structure in Usage-Based Construction Grammar
The hypothesis of the lexical origin of constructional meaning accounts for how
argument realization patterns come to be associated with meaning in a usage-based
model of grammar: as suggested by the evidence for the close relation between the
meaning of constructions and their verbal distribution, constructional meaning
could emerge from a frequent co-occurrence of a lexical meaning with a syntactic
form. However, while this relation has been documented for many constructions,
it might not always be as straightforward as previously suggested.
Insofar as the lexical origin hypothesis states that the meaning of the most
typical verb or verb class in a given syntactic pattern comes to be symbolically
associated with this pattern, it requires the assumption that all constructional
meanings are lexicalized by some particular verb or verb class in the lexicon of
the language. While this assumption has been so far shown to hold for a number
of argument structure constructions, these constructions, though probably among
the most basic in the language, constitute only a fraction of a speaker’s knowledge
of argument realization. In fact, this assumption is likely to be false for argument
structure constructions with a highly abstract meaning. In this section, I report on
such a problematic case for the lexical origin hypothesis in its present formulation:
the conative construction, which does not seem to be biased towards a particular
class of verb from which its central meaning could be extracted. It is therefore not
clear how the meaning of this construction could be derived from its typical verbs.
In Section 4.2.1, I describe the conative construction and its semantic contribu-
tion. In Section 4.2.2, I review various accounts of the construction in the construc-
tion grammar literature. In Section 4.2.3, I turn to an analysis of the construction’s
usage, using a collexeme analysis of the verbs occurring in it. I find that the lexical
origin hypothesis, even in its adjusted formulation, can hardly account for the rela-
tion between constructional meaning and constructional usage, since the construc-
tion preferably attracts verbs that do not elaborate its putatively central meaning.
syntactic relation ‘object’↜” (i.e., the patient argument). However, the semantic ef-
fects associated with the conative construction can also concern other aspects of
the transitive variant than properties of the patient. In cognitive semantic analyses
of transitivity (Hopper & Thompson, 1980; Langacker, 1991; Taylor, 1995), the
prototypical transitive event is claimed to have the following properties:
Missed contact
A first possible semantic effect of the conative is to prevent the entailment that
contact is made between agent and patient (Broccias, 2001; Dixon, 1991; Goldberg,
1995; Levin, 1993; Schlesinger, 1995). The action of the agent is directed towards
the patient, but the use of the verb in the conative construction leaves it unspecified
whether contact is actually made. Example (1) below presents such a use.
(1) He hit at her face with the gun, but she jerked her head back (BNC H85-837)
The concessive clause introduced by but makes it clear that the agent’s action described
by the conative clause is unsuccessful, because the patient changed location before it
could be reached. The transitive counterpart (he hit her face with the gun) would result
in a semantic incongruity, since transitive uses of hit entail contact with the patient.
Lack of affectedness
Second, some conative uses lack the entailment that the patient is affected by the
agent’s action, in contrast to the transitive counterpart (Broccias, 2001; Dixon,
1991; Gawron, 1983; Goldberg, 1995; Guerrero Medina, 2011; Levin, 1993; Pinker,
1989; Schlesinger, 1995). Contact is made but does not bring about a perceptible
effect on the patient. This is evidenced in Example (2) below.
92 Argument Structure in Usage-Based Construction Grammar
The sentence describes an event in which the agent tries to remove tears that have
flowed on her cheeks by wiping them off. The adverb ineffectually specifies that her
attempt is unsuccessful. The transitive counterpart (wiping ineffectually the wet-
ness on her cheeks) would sound less acceptable because it would strongly imply
that the action of the agent have a significant effect on the patient. Acceptability is
improved if the adverb is removed.
Lack of completion
The conative sentence can also differ from its transitive counterpart in that the
former lacks completion; as it were, it refers to a partly carried out action (Broccias,
2001; Gawron, 1983; Van der Leek, 1996). This lack of completion makes the event
repeatable, which is evidenced by the presence of the adverb again in Example (3):
(3) He gulped at the beer again (BNC GWF-1031)
The transitive counterpart he gulped the beer again would sound odd, because the
presence of again entails that the same event occurred before. This is however not
possible, because the transitive use of gulp entails a total ingestion of the liquid
referred to by the object. Since the definite article forces the direct object to refer
to a unique entity, the same event of total ingestion referred to by the transitive
counterpart cannot occur more than once (compare with he gulped one beer again).
The relevant interpretation of the conative counterpart is that only part of the
liquid has been drunk, allowing the event to be repeated.
Repetition
Somewhat related to the lack of completion, another relevant contrast between
conative and transitive sentences is that the conative counterpart is no longer punc-
tual (Broccias, 2001; Guerrero Medina, 2011). Contrary to the transitive, it does
not refer to a single, clearly delimited event, but rather to an undefined number
of repetitions of the same event (a multiplex, in Talmy’s [1985] terminology). In
Example (4) below, the adverbial clause missing repeatedly entails that the event
described by the conative clause occurs more than once: some instances of this
event result in successful contact, others in failure.
(4) Macready struck at Carradine’s dancing legs, missing repeatedly
(BNC GVL-1399)
The transitive counterpart Macready struck Carradine’s legs would be much less
acceptable. It would clash with the adverbial clause, because the transitive use of
strike is punctual: it refers to a single event and is therefore incongruent with an
Chapter 4.╇ The usage basis of constructional meaning 93
Lack of intentionality
Another relevant semantic aspect of the conative construction is that it can de-
scribe events in which the agent does not act with the intention of bringing about
an effect on the patient (Dixon, 1991; Guerrero Medina, 2011; Van der Leek, 1996).
This is the relevant interpretation in Example (5) below:
(5) They wandered on, aimlessly kicking at the pine cones (BNC B3J-2291)
In (5), the adverb aimlessly explicitly specifies that the agent’s action does not have
a particular goal. Again, the transitive counterpart would be less felicitous than
the conative version, as it would entail that affectedness is sought by the agent. In
such uses, the reason why the conative construction is preferred over the transitive
construction is usually that the motivation behind the agent’s action is other than
the intention to bring about an effect on the patient. Rather, the agent acts out of,
e.g., anger (Example [6]), nervousness (Example [7]), anxiety (Example [8]), panic
(Example [9]), playfulness or idleness (Example [5]).
(6) She kicked at the chair-leg and the black, angry words burst out of her.
(BNC AC4-254)
(7) She moved restlessly and kicked at the crisp leaves at her feet. (BNC HWE-682)
(8) Robyn bit worriedly at her lip. (BNC HGT-3558)
(9) She bit at her lower lip, stifling a feeling of rising panic. (BNC JXW-1811)
In these cases, if any contact with the patient or effect on it actually occurs, it is
purely contingent and irrelevant; in other words, it is not the focus of attention
(Dixon, 1991). This is clearly exemplified by (10) below:
(10) Albert closed his eyes and moaned, ‘Oh, I hate myself!’ Then he opened them
and kicked savagely at a lump of dried mud, sending it sailing across the yard
to explode against the side of the stable. (BNC EFJ-2876)
In this example, the act of kicking is motivated by the agent’s anger. Nevertheless,
the action of the agent has an objectively observable effect: that of sending the
lump of mud away. But this effect was not the intention of the agent: it is merely
concomitant with the agent’s action; incidental, as it were.
This use of the conative is also relevant to cases in which the agent is inani-
mate (Broccias, 2001); e.g., natural forces or abstract entities, as in Examples (11)
to (13) below. Since the agent is not volitional, it cannot be ascribed intentions or
94 Argument Structure in Usage-Based Construction Grammar
In the next section, I discuss how the semantics of the conative construction can
be captured in terms of a constructional meaning, along the lines of construction
grammar.
As it should already be clear from the previous discussion, the semantic contribu-
tion of the conative construction is highly variable, and, if anything, difficult to
grasp with a single generalization. What could stand as the common motivation
behind all these uses is the very abstract notion that the conative construction
moves the focus on what the agent is doing, regardless of whatever effect this ac-
tion brings about. This proposal echoes Dixon’s (1991) analysis, who notes that
“the emphasis is not on the effect of the activity on some specific object […] but
rather on the subject’s engaging in the activity” (p.â•›280). While this account seems
reasonable at first sight, such an abstract characterization must still go a long way
towards the actual semantic contribution with individual verbs, leaving a heavy
burden to processes of meaning construction. In addition, such a general mean-
ing could not account for why some verbs (such as break and bend) cannot occur
in this construction, since a priori any verb meaning involving an agent could, in
theory, undergo a focus on the agent’s activity.
Some analysts have attempted to subsume the various uses of the conative con-
struction under the meaning ‘directed action’ (cf. Pinker, 1989, p.â•›104; Goldberg,
1995, p.â•›63). This analysis allows them to unify the conative construction with
another use of the syntactic pattern [Subject Verb at NP], which may be called
the directed-action construction, in which the at-phrase clearly refers to a target
towards which an agent directs some action, as with verbs of looking, verbs of
sound emission or verbs denoting facial expressions (inter alia), as illustrated by
Examples (14) to (16).
(14) Jim looked/stared/glanced at the flyer.
(15) She was shouting/screaming/laughing/sniggering at her husband.
(16) The child smile/smirked/winced at him.
Chapter 4.╇ The usage basis of constructional meaning 95
However, such an account is problematic with regards to the productivity and the
semantic contribution of the construction. For verbs to be compatible with this
constructional meaning, their semantics must provide an “orientational” compo-
nent, i.e., an entity that can be oriented in some direction (cf. Perek & Lemmens,
2010, §27). Verbs of striking (such as kick, punch, jab) plausibly do, as well as the
semantically similar verbs of thrusting (e.g., thrust, poke, prod). With such verbs,
the orientational component may be identified with the body part or instrument
oriented (and moved) by the agent in the direction of a target: a foot for kick, a fist
for punch and pummel, a finger for poke, a pointed object for jab and prod, etc. Verbs
of seizing (clutch, grab, snatch) could also possibly be accommodated into this con-
structional meaning by identifying the orientational component with the moving
hand of the agent. For these three classes, since the agent’s action is purely direc-
tional, the conative use of these verbs leaves unspecified whether contact is actually
made, which corresponds to the first semantic effect I identified in Section 4.2.1.
However, the meaning ‘directed action’ loses much of its relevance if we con-
sider other verbs whose inherent directionality is less obvious, if it exists at all.
While for verbs of pushing (push, shove, nudge) the orientational component could
be assumed to be the force itself that the agent exerts on the patient, verbs of pull-
ing (pull, tug, yank) are more problematic: while they also refer to the exertion of
a force, this force is not oriented towards the patient but in the opposite direction;
hence they violate the semantic specifications of the directed-action construction.
For verbs of rubbing (rub, brush, scrape), the action of the agent is not directed
towards the patient but rather exerted all over it; in addition, contact is made with
the patient, even in the conative uses, which means that the unspecified contact
interpretation contributed by the construction is not relevant. In the case of verbs
of ingestion (sip, nibble, gulp) and biting (bite, chew, gnaw), the orientational com-
ponent is, at any rate, of a very peculiar kind.
It might be argued that for verbs whose semantics do not straightforwardly fit
into the ‘directed action’ meaning, the directionality implied by the construction
is understood abstractly. For example, with verbs of pulling, it could be assumed
that the force is directed towards the patient in an abstract sense, in that it is a force
that the patient is subjected to, although it is not directed towards it in a concrete
sense. A similar analysis may hold for other problematic verb classes. At any rate,
the abstract construal of directionality does, however, also not entirely account
for the full range of semantic effects observed with the conative construction. As
mentioned earlier, the only semantic effect that the directed-action construction
could contribute by itself is to leave contact unspecified. The lack of affectedness
could possibly be integrated by transposing the ‘directed action’ event from the
physical field to the causal field (Jackendoff, 1983), where the agent in a causative
event is understood as an entity moving towards a goal (bringing about of an effect
96 Argument Structure in Usage-Based Construction Grammar
on the patient), which is not reached in the conative variant. But the other seman-
tic aspects of the construction (completion, repetition and lack of intentionality)
remain largely unexplained (or only poorly so) under the assumption of a unique
‘directed action’ meaning. Such problematic data indicate that a lumping strategy
of subsuming the conative construction under the directed-action construction is
highly questionable. Broccias (2001) raises similar concerns about a monosemic
analysis of the conative construction, which he instead associates with three differ-
ent constructional meanings that are only partly related to the meaning ‘directed
action’, if at all: (i) the allative schema, “where the action denoted by the (transitive
use) of the verb does not necessarily take place” (p.â•›69), (ii) the ablative schema
“where the verbal event does take place, though in a bit-by-bit fashion” (idem) and
(iii) the allative/ablative schema, which combines aspects of the first two schemas.
The allative schema describes the emission of a force (which may be metaphor-
ically construed) towards a goal, but with forceful contact not being linguistically
encoded and thus backgrounded and merely implied. This description appears to
be similar to the meaning of the directed-action construction, but note that the
allative schema and the direction-action construction differ in that the former
includes both a force component and a motion component, whereas the latter
only refers to an orientational component, from which the directional component
is merely implied. In addition, the directed-action construction does not involve
actual motion: the directional component is merely a vector projected by the ori-
entational component towards the referent of the at-phrase. The allative schema is
exemplified by sentences (17) and (18) below. In (17), the emitted entity is a foot,
as lexicalized by the verb kick; in (18), it is a hand, as lexicalized by clutch. The al-
lative schema is compatible with many other verbs of striking (such as hit, smash,
jab) and seizing (such as grab and snatch).
(17) Sally kicked at the wall. (from Broccias, 2001, p.â•›73)
(18) He clutched desperately at the branch as he fell. (idem)
The ablative schema is associated with continuous actions, either because it is re-
peated (i.e., in a “bit-by-bit” manner) or because one single instance of the action
is prolonged. This schema bears some connection to the allative schema, in that
it incorporates a notion of attempt, which, however, does not apply to the verbal
event itself but rather to its expected or desired consequences. In other words,
some action is performed upon an entity in an attempt to bring about some result,
which, according to Broccias, can belong to one of the following three categories:
(i) removal of a substance or an entity (e.g., caused motion of an entity as in [19]),
(ii) release (the coming about of a perceptual state predicated of either the subject
referent or the object referent, e.g., sensation in [20], attention in [21] and informa-
tion in [22]), or (iii) creation/destruction (e.g., ingestion in [23]).
Chapter 4.╇ The usage basis of constructional meaning 97
(19) The horse pulled at the cart. (from Broccias, 2001, p.â•›75)
(20) … and tears stung at her eyes. (p.â•›77; from BNC HGK-2219)
(21) The child pulled at his mother’s coat, wanting to be lifted up.
(p.â•›78; from LDELC dictionary)
(22) Cardiff nudged at that swollen head with his shoe. (p.â•›78; from BNC G0E-3258)
(23) James Bond sipped at his Martini.
Example (21) contrasts with its transitive counterpart not by entailing that the action
of pulling may not have been carried out, but rather by leaving unspecified the conse-
quence of that action. Similarly, (23) contrasts with its transitive counterpart (James
Bond sipped his Martini) in that it entails that the drink has not been fully ingested;
it also contrasts with examples of the allative schema in that it does not feature the
same notion of attempt, as some of the Martini has indeed been successfully drunk.
As previously mentioned, some uses instantiate the allative/ablative schema,
which combines the two previously discussed schemas. For example, sentences
(24) and (25) below involve both the emission of a force (which may be metaphori-
cal) directed towards a target, and a continuous action in an attempt to bring about
some (future) result, not necessarily successfully.
(24) Sam sprayed at the trees with insecticide. (from Broccias, 2001, p.â•›79)
(25) He was working at his painting. (idem)
It should be noted that many instances of the allative schema may also be catego-
rized as instances of the directed-action construction. This happens most likely
with verbs of striking, verbs of thrusting, and possibly verbs of seizing, which
provide an orientational component. In other words, there is likely to be some
overlap between the two constructions, which is probably why some researchers
have suggested that they might be one and the same. Furthermore, both construc-
tions probably share a common origin in diachrony related to an archaic meaning
of the preposition at, which may add to the confusion between them. At is the
contemporary version of Old English æt (related to Old Norse and Gothic at, Old
Frisian et and Old High German az), which can be traced back to a Proto-Indo-
European root, *ad-, meaning ‘to’ (cf. Latin ad ‘to, toward’). In present-day English,
at basically means “in the close neighborhood of ” (e.g., at home, at the door, at
the swimming pool, etc.). The earlier directional sense has plausibly been frozen in
some verbal constructions of English (and their corresponding nominalizations),
but the preposition no longer conveys a notion of directionality on its own (for
example, the path at the beach cannot mean ‘the path towards/to the beach’).
Let us return to the several abstract schemas that the conative construction
more likely corresponds to. Whether or not these schemas can be related in a
polysemic network is a matter of debate, but it appears to be a reasonable position.
98 Argument Structure in Usage-Based Construction Grammar
Having examined more closely the meaning of the conative construction, I now
address the question of how this meaning can be derived from the meaning of the
verbs typically occurring in it. The verbal distribution of the conative construc-
tion was extracted from the prose-fiction part of the British National Corpus,
containing about 16 million words in 431 texts primarily drawn from novels. The
choice of this corpus was neither arbitrary nor unmotivated. Intuitively, the cona-
tive construction seems to carry a complex descriptive function which makes it
particularly useful for narrative purposes (especially in written discourse). While
I argued in Chapter 3 that spoken language (and especially conversation) should
be preferred in studies of linguistic usage, since it is arguably the most basic genre
and represents a major part of a speaker’s linguistic experience, a corpus of novels
seems more appropriate for the study of the conative construction as it is used in
the production of English writers. The intuition that the conative construction is
probably not to be found so frequently in spontaneous spoken language was borne
out by an earlier attempt at finding conative sentences in the conversation part of
the corpus, revealing that the construction is extremely rare in that register (only
17 tokens in 4 million words, i.e., 4.25 per million words). It is indeed much more
common in the prose-fiction corpus, with 160.19 instances per million words.
I queried the corpus for all verbs followed by the preposition at, with an op-
tional intervening adverb. I excluded frequent verbs that cannot support a conative
reading and for which at can only be used in a purely locative sense (e.g., be, stay,
Chapter 4.╇ The usage basis of constructional meaning 99
live, arrive, etc.). The resulting set of sentences was manually annotated to select
only conative sentences, which were defined according to two criteria: (1) the verb
has to be transitive and (2) the interpretation of the sentences has to fall into one
of those described in the previous section. Sentences with coordinated verbs were
duplicated in the dataset (one duplicate per verb). This yielded a final set of 2,563
instances, distributed over 159 verb types.
5. Available at http://www.linguistics.ucsb.edu/faculty/stgries/teaching/groningen/
6. As previously mentioned, a collexeme is significantly attracted (at the 5% level) if its col-
lostruction strength is above 1.301.
100 Argument Structure in Usage-Based Construction Grammar
7. It is, however, not clear whether they did so on the basis of verb meanings or referential
information alone, since they simultaneously learned the nonce construction and the nonce
verbs occurring in it.
Chapter 4.╇ The usage basis of constructional meaning 101
verbs, each with relatively low frequency” (p.â•›304). In fact, Perek and Lemmens
(2010) argue that the meaning ‘directed action’ associated with the syntactic pat-
tern [Subject Verb at NP] when it is instantiated with intransitive verbs (which, as
argued in Section 4.2.2, must be considered a distinct generalization from the co-
native construction) may well be derived through such a process. They report that
the directed-action construction is strongly biased towards verbs of looking (espe-
cially look). Hence, the meaning of the construction is at odds with the meaning
of its most typical verbs, since the former is more abstract than the latter. Yet, the
central meaning of the construction can arguably be derived by abstracting away
the semantic commonality among verbs of looking and other prominent verbs in
its distribution, namely the orientational component mentioned in Section 4.2.2.
In the previous section, I argued that there are good reasons to consider what
Broccias (2001) termed the allative schema as the central meaning of the cona-
tive construction, defined as the emission of a force towards a goal. The other
semantic aspects of the construction can arguably be derived from this central
meaning through polysemic extensions. Hence, the emergence of the semantics of
the conative construction could amount to the abstraction of the allative schema
from its typical verbs. For this to be plausible, the construction has to attract many
verbs which make reference to a force emitted along an unbounded trajectory. As
mentioned in the previous section, verbs of striking and verbs of thrusting (listed
together among “verbs of hitting and touching” in the list above) arguably feature
this semantic component. Verbs of seizing also possibly do, but they are clearly less
typically directional in nature, since they profile the action of taking hold of some-
thing rather than the movement which led to this action. As it turns out, directional
verbs do not appear particularly high in the list. The most prominent collexeme,
tug, instantiates the ablative schema, like all other verbs of pulling; as previously
argued, it does not contain a directional element stricto sensu. Thus, an attempt at
deriving the central meaning of the construction from this verb would clearly be
misguided. The verbs clutch8 and dab do occur respectively in second and third
position, but they do not so strikingly feature the “emission of a force”: the former
is a verb of seizing and while the latter can be considered a verb of striking, it does
not profile a particularly strong force component. Besides, they are the only two
verbs of this kind among the very top collexemes: the next representative of this
class, stab, is only ranked 13. All in all, verbs whose inherent meaning most clearly
8. It must also be noted that clutch and other verbs of seizing frequently occur in the construc-
tion in their static ‘holding’ sense. The exact ratio of seizing vs. holding uses is hard to evalu-
ate, since many instances in context are ambiguous between the two interpretations. From an
observation of corpus examples, the holding sense seems quite frequent. At any rate, the two
senses would probably appear lower down the list if they were treated as distinct collexemes.
102 Argument Structure in Usage-Based Construction Grammar
elaborate the allative schema (jab, slash, swipe, poke, prod and kick) are clustered in
the second half of the list and their collostruction strength is much lower than that
of the collexemes of the top half of the list. By way of contrast, verbs instantiating
the ablative schema or the allative/ablative schema are much better represented.
Besides tug, other prominent collexemes which arguably do not instantiate the
allative schema include gnaw, nibble, sip, pluck and hack.
In sum, contrary to what was found with a number of other constructions,
collexeme analysis is not helpful in identifying a particular sense of the conative
construction which would be central and from which the other senses could be
derived. The allative schema, described by Broccias (2001) as profiling the emission
of a force towards a goal, is plausibly the central meaning of the construction, ac-
cording to the analysis presented in the previous section. It could be assumed that
this abstract meaning is generalized from the semantics of several distinct verbs,
without one verb lexicalizing the constructional meaning as such (which cannot
be the case, since such a verb does not exist). If that were the case, the construction
would attract first and foremost verbs whose semantics profile such a directional
component, and only secondly verbs of other kinds. What was found is exactly
the opposite: the construction primarily attracts verbs which are not inherently
directional or only loosely so, and verbs whose directionality is more salient are
less strongly attracted. This is in striking contrast with the examples presented in
Section 4.1.4, where the most representative instantiations of the central meaning
of the construction were among the very top collexemes.
In conclusion, the case of the conative construction conflicts with the lexical
origin hypothesis, in that there is a mismatch between the verbs that the construc-
tion most strongly attracts and the central meaning of the construction, which can
only be derived from verbs which figure less prominently among the construction’s
collexemes. This case study shows that the lexical origin hypothesis in its present
formulation is an incomplete proposal. It does not tell the whole story about the
emergence of constructional meaning: while it provides an explanation for the
facilitating effect of verb biases and makes accurate predictions for many construc-
tions, it is by itself insufficient.
4.3â•… Conclusion
As recent research seems to indicate, there is substantial evidence for the usage
basis of constructional meaning in the domain of argument structure. Many stud-
ies point to a close connection between the meaning of an argument structure
construction and the meaning of verbs that typically occur in it. This finding,
supported both by the study of constructional usage and experimental evidence,
Chapter 4.╇ The usage basis of constructional meaning 103
In this chapter, it is argued that the conflicting evidence found with respect to
the lexical origin of constructional meaning stems in great part from the fact that
studies investigating the usage basis of argument structure constructions usually
consider only the highest level of constructional generalization. Instead, I suggest
that such investigations should start at lower levels of abstraction, and I show that
the relation between constructional meaning and verbal distribution can indeed
be restored at these lower levels.
The chapter is structured as follows. In Section 5.1, I discuss the concept of
schematization, whereby generalizations can be defined at several levels of abstrac-
tion in a usage-based approach. I present evidence from previous research that
lower-levels generalizations might on balance play a more important role in lan-
guage use and language development than fully abstract schemas. In Section 5.2, I
suggest that the conative construction more plausibly consists of a set of low-level
schemas instantiated by semantically delimited verb classes rather than a single
highly abstract generalization, and I submit four verb-class-specific constructions
to a collexeme analysis. I find that each of these lower-level constructions attracts
verbs whose meaning provides an indication of the semantic contribution of the
conative argument structure for verbs of the corresponding class. I conclude that
the lexical origin hypothesis does hold for the conative construction, albeit at a
more local level of generalization.
(idem). This allows the constructional schemas abstracted from usage to be defined
at any level of specificity. A schema can be abstracted from highly similar expres-
sions, thus constituting a generalization of a limited scope. If more dissimilar ex-
pressions are considered, then to the extent that these expressions still share some
commonality, a more abstract schema may emerge, constituting a generalization
of a broader scope. For Langacker, “linguistic patterns occupy the entire spectrum
ranging from the wholly idiosyncratic to the maximally general” (p.â•›3).
This concept can be illustrated with an example involving transitive construc-
tions. Let us assume a speaker who already knows the following schemas capturing
the transitive use of the verbs curl, fold, stab and whip:
(1) a. [ np curl np ]
b. [ np fold np ]
c. [ np stab np ]
d. [ np whip np ]
These schemas already constitute abstractions over a particular use of these four
verbs that are less specific than actual instances: they specify a lexical form (though
no tense or aspect marking), but leave open slots for the subject and object comple-
ments. The representations sketched in (1) lack a semantic pole; the latter would
contain the schematic meaning of the verb, including semantic information about
the kind of entities that can be filled into the subject and object slots.
From these four schemas, more abstract ones can be derived on the basis of
formal and semantic similarities. Curl and fold both refer to an event in which some
entity (the agent, realized in subject position) acts on some other entity (the pa-
tient, realized in object position) by physical means such that the latter undergoes
a change in its overall shape. Hence both schemas are semantically very similar;
they only differ as to the kind of resulting shape that the patient finds itself turned
into: a curve in the case of curl, and for fold a spatial configuration in which one
part of the folded entity lies on another part (both parts are typically flat surfaces).
A more abstract schema can thus be posited over them that retains their semantic
and syntactic commonalities. Of course the entities that can be curled differ from
those that can be folded, but the ‘cause change of shape’ schema blurs this differ-
ence. A similar process can occur with the transitive uses of stab and whip. They
both refer to an event in which an agent acts in a specific way on a patient (push-
ing a pointed object into it for stab and hitting it with a whip or another flailing
weapon for whip) so as to cause it harm; they only differ with respect to the means
employed by the agent (and correspondingly, the kind of injury suffered by the pa-
tient). On the basis of these commonalities, a schematic meaning ‘cause harm’ can
be posited. The ‘cause change of shape’ and ‘cause harm’ schemas are represented
in Figure 5.1, along with the schemas they abstract over. These schemas constitute
the first level of abstraction in the fictitious grammar outlined in (1).
Chapter 5.╇ The importance of local generalizations 107
NP VERB NP
‘cause effect’
SHAPE HARM
NP VERB NP NP VERB NP
‘cause change of shape’ ‘cause harm’
As Figure 5.1 indicates, an even higher level of abstraction is possible. The ‘cause
change of shape and ‘cause harm’ schemas both contain the notion that the agent
acts on the patient so as to bring about an effect on it. A higher-level schema can
thus be abstracted over these two intermediate ones, retaining this common mean-
ing, as shown in Figure 5.1.
Of course, this simple example, involving only four verbs, does not by far ex-
haust the various abstractions that can be made over the many transitive verbs of
English, and other intermediate-level schemas could be posited for verbs which do
not fall into the generalizations ‘cause change of shape’ or ‘cause harm’. Also, other
transitive expressions could support an even more abstract generalization than the
‘cause effect’ schema, in which the ‘effect’ component would be absent and only
the notion that the agent acts on the patient would remain. That being said, this
example illustrates what is meant by “degree of granularity”: the transitive uses of
curl and fold seem different when taken in all of their semantic details, but they
turn out to be similar to some extent when some of these details are blurred over,
and the similarity between curl and stab appears at an even higher level of abstrac-
tion. As Langacker (2000, p.â•›25) puts it, “there is no inherent limit to the level of
specificity at which a constructional schema can characterize its components”.
This theoretical claim does however not decide on exactly which generalizations
speakers actually make, and which are most relevant to language in use. I turn to
these questions in the next section.
108 Argument Structure in Usage-Based Construction Grammar
study of ten children, early child speech contains a mixture of item-based and
more abstract patterns. Hence, low-level schemas can be seen as pathways towards
adult grammatical competence: they are the building blocks from which abstract
argument structure constructions are derived. Returning to my initial question, if
low-level schemas are formed early in language development, it is hard to imagine
that children would abandon them once higher-level schemas have been acquired.
Although adult grammar is less markedly item-specific than child grammar, the
developmental data suggest that it might well still contain generalizations of vary-
ing scope. Hence, a cognitively sound approach to grammar should first focus on
lower levels of abstraction, before positing broader generalizations, as also argued
by Iwata (2008):
[I]f one strictly follows the idea of a usage-based theory, children learn a language
by inductively generalizing over usage-events. Verb-class-specific constructions
are nothing more than generalizations at a relatively low level, so a large number
of lower-level constructions being stored in long-term memory coincides with
such fundamental linguistic behavior, and so should be seen as natural. (p.â•›212)
One factor that could lead to the decay of constructional generalizations (at any
level of specificity) is their lack of reinforcement, i.e., if a conceptual structure (of
any kind) is not activated for a long time, it tends to fade away from long-term
memory. However, according to Langacker (2000), such a scenario is not likely to
happen for many low-level schemas. In Langacker’s model, a usage event, either of
production or comprehension, amounts to finding a “standard” for the categoriza-
tion of a “target” (the message to be conveyed in production, or the expression to
be interpreted in comprehension). In any given situation, there may be more than
one available standard for categorization: this is the activation set. Several factors
come into play in the choice of which particular member of the activation set wins
the competition and emerges as the active structure to categorize the target: (i) the
level of entrenchment, (ii) contextual priming (i.e., which structure is the most
relevant given the context of utterance) and (iii) the amount of overlap between
the target and the available standards. This latter factor is relevant to low-level
schemas: the more specific a schema is, the more similar it is to actual utterances,
and the more likely it will win the competition, particularly over more abstract
categorizing structures. As Langacker (2000) puts it:
We can reasonably assume that the sharing of features is what enables the target
to stimulate members of the activation set in the first place, and that the degree
of stimulation is roughly proportional to the number of features shared. This has
the important consequence that lower-level schemas, i.e. structures with greater
specificity, have a built-in advantage in the competition with respect to higher-
level schemas. (p.â•›16)
Chapter 5.╇ The importance of local generalizations 111
Note that since abstract schemas are immanent in their elaborations, they are
activated along with lower-level schemas, although to a lesser extent. At any rate,
Langacker’s model, defined in accordance with what is known about human cog-
nitive processing (cf. Langacker, 2000, Section 3), predicts that low-level schemas
are more frequently activated than higher-level ones by virtue of their conceptual
advantage for the categorization of novel expressions. For that reason, they are not
likely to decay through lack of reinforcement.
Grammatical patterns can be defined at any level of generality: there can be highly
abstract constructions that license the argument structure of a high number of
semantically varied verbs, and there can be more local schemas licensing the same
argument structure for more semantically specific verbs. Both types of general-
izations can co-exist in grammar, but there is evidence that lower-level schemas
may be more important in a cognitively realistic account. Thus, the study of the
grammar of verbs should begin at the lowest levels of generalization if it is to
achieve any level of cognitive realism. Yet, these levels have been largely ignored in
investigations of the relation between constructional meaning and lexical meaning.
Collostructional studies often start with a formally defined syntactic pattern and
investigate the verbs typically occurring in that pattern.
I would suggest that this is one of the reasons why investigations of the lexi-
cal origin hypothesis give conflicting results for different constructions. They try
to attest the relation between constructional meaning and lexical meaning at the
highest level of generality, but if a construction may be defined at any level of gen-
erality, it might be the case that the relation between constructional meaning and
verbal distribution holds at some levels but not at others. If the lower-level schemas
are more relevant to the definition of constructional meaning, evidence for the
lexical origin hypothesis must be sought at these levels before the hypothesis can
be discarded. In the remainder of this chapter, I apply this proposal to the study
of the conative construction.
To study the relation between constructional and lexical meaning at lower levels of
abstraction in the conative construction, I draw on a proposal by Croft (2003), who
suggests, as an alternative to the notion of constructional polysemy, that argument
112 Argument Structure in Usage-Based Construction Grammar
In other words, the various senses of the ditransitive construction are not equally
available in any given context. In fact, only one of these senses is ever active in any use
of the construction, which depends on the verb it is instantiated with (see Kay, 2005
for a similar point). By way of contrast, a polysemous word like position can receive
different competing (although possibly related) interpretations in a given sentence.
For example, in the sentence I found a good position, the noun position can be in-
terpreted as either ‘job’, ‘spot’ or ‘spatial configuration of the body’. Without further
context, all three interpretations are available. Hence, the ditransitive construction
should not be considered truly polysemous in the same sense as lexical items.
Croft concludes that “a proper representation of the construction schema for
each ‘sense’ of the ditransitive construction must specify which verb classes occur
Chapter 5.╇ The importance of local generalizations 113
with it” (p.â•›56). Hence, instead of a highly abstract pattern associated with differ-
ent meanings, the ditransitive construction is better described as a collection of
what Croft terms “verb-class-specific-constructions”, i.e., lower-level generaliza-
tions of a constructional meaning over a clearly delimited semantic verb class,
instantiated only with verbs of that class. The meaning of each verb-class-specific
constructions includes the relevant modulation of the central ‘transfer of posses-
sion’ meaning, which translates into specific semantic restrictions on the verb slot
corresponding to the particular class they are associated with. Thus, Croft’s study
can be added to the list of arguments in favor of lower-level generalizations that
I presented in the previous section.
A polysemic analysis of the conative construction runs into similar problems to
those reported by Croft for the ditransitive construction. While there can be several
different readings of a single conative sentence, not all interpretations are equally
available in all instances. If Broccias’ (2001) allative schema and ablative schema
are taken to be two established senses of the construction (cf. Section 4.2.2), it
appears that each sense corresponds to different classes of verbs. As mentioned in
the previous chapter, the allative schema is instantiated only by verbs of striking,
verbs of thrusting, verbs of seizing, and possibly other verbs which can be assimi-
lated to either of these classes. For example, in no case would conative sentences
with verbs of ingestion be interpreted with the allative schema, thus conveying the
meaning ‘X moves towards Y in order to ingest Y’. Conversely, the ablative schema
is associated with another range of semantic classes: verbs of ingestion (sip, nibble,
gulp), verbs of pulling (pull, tug, haul), verbs of pushing (push, shove, nudge), verbs
of biting/chewing (bite, chew, gnaw), verbs of cutting (cut, chop, snip), verbs of
rubbing (scrup, sweep, wipe), and many other smaller classes.
It must be acknowledged that there is certainly more overlap between the
distributions of the allative and ablative schema than between those of the senses
of the ditransitive construction. This is made clear by Broccias’ (2001) positing the
allative/ablative schema, which by definition is instantiated by verbs compatible
with both the allative and the ablative schema. Despite this overlap, many verbs
are fairly restricted to only one of the two schemas. Sometimes the unavailability
of some reading is straightforwardly explained by intrinsic properties of the verbs
themselves: for example, the impossibility of an incremental reading with semel-
factives such as hit and kick can be explained by the aspectual properties of these
verbs and more particularly the absence of an incremental theme. But there are still
perfectly sensible combinations which nonetheless are disallowed, which would
not be the case if the construction was truly polysemous. There are thus good
reasons to assume that the conative construction would be more appropriately
described as a set of verb-class-specific constructions.
114 Argument Structure in Usage-Based Construction Grammar
1. Note that I am explicitly dealing with the particular problem of determining semantic classes
of verbs. This contrasts with the more general problem of grouping verbs according to their
linguistic properties, which has received a lot of attention in the literature (e.g., Dixon, 1991;
Faber & Mairal Usón, 1999; Levin, 1993).
116 Argument Structure in Usage-Based Construction Grammar
to several meanings, and these meanings can belong to different semantic classes.
For example, peck can function as a verb of eating but also as a verb of striking. The
first step thus consists in annotating the corpus data with verb senses.
The semantic information necessary to annotate the meaning of verb forms
can be drawn from any dictionary. However, the present approach requires more
than just distinguishing verb senses: it also calls for information about how the verb
senses of the distribution are related, and more particularly whether they share a
common hyperonym. Hyperonymy can be determined on an intuitive basis, but it
would surely make the task easier and the results more reliable if this information
was already provided by the same lexicographic resource. It is chiefly for this rea-
son that I used the WordNet database (Miller, 1995; Fellbaum, 1998). WordNet is
a large lexical database of English that groups words into sets of synonyms (called
synsets) and provides lists of the various meanings of each word form, which can
be looked up to perform semantic annotation. WordNet records relations between
synsets such as hyponymy, hyperonymy, part-whole relations, entailment, etc. Of
particular interest for this study, the relations of hyponymy (and conversely, hy-
peronymy) connect the synsets into a type hierarchy, which can be used to define
verb classes, in line with the proposals of the previous section. In WordNet terms,
a verb class includes the verbs of a given synset and all of its hyponyms, i.e., verbs
whose meaning includes (and often, elaborates) the meaning of the synset. Hence,
co-hyponyms belong to the same class. In sum, WordNet can be used both to an-
notate for verb senses and to define verb classes on the basis of the annotated data
and hyponymy/hyperonymy relations between senses recorded in the database.
Another benefit of WordNet is that it is freely available in electronic form, which
enables computer-assisted analysis.
At this point, a few comments are in order regarding my choice of WordNet
as opposed to another popular resource, FrameNet. The FrameNet project is the
lexicographic implementation of Frame Semantics. Its goal is to build a lexical da-
tabase of English in which the description of lexical items is based on the concept of
semantic frame. As of July 2014, FrameNet contains 1,180 frames evoked by 12,849
lexical units (including 4,982 verbs). The database also includes a wealth of other
information such as frame elements, relations between frames (somewhat compa-
rable to the relations between synsets found in WordNet), and annotated corpus
examples (including valency realization information). Since, in the first part of this
book, I endorsed the frame-semantic approach to verb meaning, it might seem
surprising that I do not explicitly apply the same approach to determine semantic
verb classes, by using FrameNet to annotate the distribution of the conative con-
struction with semantic information. In fact, I did try to use FrameNet for semantic
annotation, but this turned out not to be feasible for two reasons. First, despite the
impressive statistics of the database reported above, many lexical units occurring
Chapter 5.╇ The importance of local generalizations 119
in the conative construction are still missing in FrameNet; examples include bite,
catch (in the sense of ‘seize, grasp’), hack, suck (in the ingestion sense), and wipe.
Hence, when I attempted to annotate the conative dataset for frames using the
FrameNet 1.5 data,2 982 tokens (37%) were left without annotation because of
missing information in the database, versus 34 tokens (0.01%) with WordNet.
Admittedly, at least some of these gaps could be easily fixed by attaching missing
lexical units to existing frames (but not so easily if new frames have to be created).
At any rate, the necessity to complement FrameNet with the lexical information
it is lacking defeats the purpose of using an external resource in the first place.
Besides, adding my own data to FrameNet would also lead me to make decisions
that could potentially be inconsistent with the rest of the database, and, more
importantly, it would jeopardize overt operationalization (cf. Glynn, 2010), in the
sense that the analytical criteria of the study would no longer be overtly identified
in such a “customized” version of FrameNet (as opposed to the public version).
Second, another problem is that even for the lexical units that are described
in the database, the FrameNet frames often do not encode fine-grained enough
distinctions. Many of the frames are too general and include a very wide diversity
of verbs; as such, they clearly do not correspond to the basic level of categorization
that I suggested as a criterion for semantic verb classes. For instance, such widely
diverse verbs as clutch, lick, nudge, pull, push, rub, squeeze, and touch are all lexi-
cal units of the “Manipulation” frame, which describes “the manipulation of an
Entity by an Agent”. Clearly, there is more to the meaning of these verbs than what
FrameNet encodes (at least in its current state). The very abstract meaning of the
Manipulation frame blurs over many semantic details that could be used to discern
lower-level classes. In sum, the semantic distinctions between FrameNet frames
are too broad to form the basis of verb-class-specific constructions of the kind
envisaged in Section 5.2.1. It is not clear if the lack of finer-grained distinctions is
intended or if it is merely a contingent situation due to the fact that FrameNet is still
under construction, and which should be fixed as the project advances. I personally
have the intuition that there is a limit to the degree of semantic distinctiveness that
FrameNet is designed to capture, although I do not believe it is a serious limita-
tion of the approach in general (though it might be for the specific purpose of this
study). Be that as it may, if finer-grained distinctions than those provided by most
FrameNet frames are needed, and if, as I am about to show, WordNet is able to
provide such a degree of granularity, there is no point in including the coarser-
grained layer of annotation that FrameNet would provide.
slowly or heavily’. The semantic difference between these two senses is, if anything,
very subtle, and therefore not easily applicable to the annotation of examples in
context. However, since both senses have pull as their direct hyperonym, they can
be conflated into a single entry subsumed by the class of verbs of pulling. Even
though the sense distinctions posited in WordNet might not always be easy to
draw, the coarser-grained distinctions imposed by verb classes are more reliable
and more easily noticeable.
The next step consists in using the WordNet hierarchy of verbs to posit semantic
classes. To achieve this, I used an iterative method, which first consists in choosing
a set of senses in the database which could be hyperonyms of several verbs. Some of
them were already in the distribution, others had to be found by browsing WordNet.
Then a computer script searched the database for hyperonymy relations between the
hyperonym candidates and the senses of the distribution, yielding semantic group-
ings of senses as output. I then searched the remaining senses (i.e., those not related
to any hyperonym) for additional hyperonym candidates and ran the script again.
I repeated this process until there were a reasonably low number of ‘orphans’ left
(i.e., verb senses that do not belong to a verb class). When I reached a stage when no
more hyperonyms could be found, I ended up with thirteen classes, accounting for
116 senses and 1,964 tokens (i.e., 79%) of the distribution. This leaves ninety-nine
independent senses, many of which are metaphorical uses of a verb belonging to a
semantic class. The thirteen semantic verb classes of the sense distribution of the
conative construction are listed in Table 5.1 with their type and token frequency.
Each class is named after its hyperonym.4 A list of verbs belonging to this class and
attested in the conative construction in the corpus is also provided. This list is not
exhaustive, since there are other verbs in WordNet that have the same hyperonym
but were not found in the conative construction.
Several classes are particularly prominent both in terms of types and tokens.
Both criteria are important for this study. High token frequency reflects the level
of entrenchment of the verb-class-specific construction, and thus increases the
4. The hyperonym of verbs of ingestion (ingest) appears to violate at least one property of basic-
level terms, since it is not highly frequent, and at any rate markedly less frequent than drink and
eat. Hence, the lower-level classes of drinking verbs and eating verbs would appear to be better
candidates for the basic level. However, the distinction between the two classes only relies on
the nature of the theme (solid vs. fluid), and a perceptual gestalt still exists for ingest regardless
of this distinction. As a matter of fact, many verbs can be used with either kind of theme. Eat
usually refers to the ingestion of solids, but a soup is eaten although it is fluid. Other verbs, like
gulp and lick, do not select for either solid or fluid substances. The distinction between eating
and drinking is thus not as lexically grounded as would be expected if they were basic-level
categories. There is therefore good reason to posit verbs of ingestion as a semantic class of its
own, as opposed to splitting them into verbs of drinking and verbs of eating.
122 Argument Structure in Usage-Based Construction Grammar
likelihood that it has unit status (Langacker, 1987). High type frequency indicates
that the construction contains many different members, which makes collexeme
analysis of that construction more likely to provide an indication of the construc-
tional meaning by contrasting the meaning of attracted members to that of others
members of the class. Among the classes that have both a particularly high type
and token frequency, I selected four that I submitted to collexeme analysis: verbs
of cutting, verbs of hitting, verbs of ingestion and verbs of pulling. In the next
section, I describe how collexeme analysis was implemented for the study of these
four verb-class-specific constructions, before turning to the analysis itself.
Table 5.1╇ The thirteen verb classes found in the conative construction, with their type
and token frequency and a list of their members.
Class Types Tokens Members
biting 4 42 bite, nibble, nip, snap
chewing 3 64 chew, gnaw, munch
cutting 10 54 chip, chisel, chop, cut, hack, nick, saw, scratch, slice, snip
hitting 26 396 bang, bash, bat, batter, beat, chop, dab, hammer, hit, jab,
lash, kick, pat, peck, pound, pummel, punch, scuff, slap,
smash, strike, swat, swipe, tap, thump, whack, whip
holding 3 57 clutch, grasp
ingestion 11 181 eat, gobble, gulp, lap, lick, nibble, peck, pick, sip, suck, swig
pulling 11 448 drag, haul, jerk, pick, pluck, pull, tug, tweak, twitch,
wrench, yank
pushing 6 74 flick, nose, nudge, prod, push, shove
rubbing 6 82 brush, paw, rub, scrape, scrub, wipe
seizing 14 333 catch, clasp, claw, clench, clutch, grab, grapple, grasp, grip,
pinch, seize, snatch, tweak, twitch
tearing 3 89 rip, shred, tear
thrusting 6 90 dig, jab, poke, prod, stab, thrust
touching 13 54 flick, knead, knuckle, lap, lick, massage, mouth, nose,
nuzzle, paw, peck (=kiss), squeeze, stroke
Total 116 1964
their different senses can belong to different verb classes, the distribution of a verb-
class-specific construction is a distribution of verb senses. Below is the contingency
table that must be computed for each verb in a given class, adapted from Table 4.3,
with labels modified accordingly:
Filling the first line of the table is unproblematic, since all the relevant frequencies
are already available. F(S ∧ C), the frequency of each verb sense in the verb-class-
specific-construction, is taken from the distribution of verb senses in the conative
construction that I used to determine the semantic verb classes. The frequency of a
verb-class-specific construction F(C) is obtained by summing the frequency of all
verb senses in the class. With these two frequencies, the frequency of other senses in
the verb-class-specific construction can be calculated: F(¬S ∧ C)â•›=â•›F(C)â•›−â•›F(S ∧ C).
The second line of the table is a less trivial issue, and to fill it, other frequencies
are needed. F(S ∧ ¬C) is the frequency of each verb sense in other constructions.
However, the corpus can only provide frequencies of verb forms, and some of
these may actually be distributed over several semantic classes. All instances of a
verb form cannot just be assigned to a single class or be counted in several classes
simultaneously, as this would be a large source of noise: it must be determined for
each token of a given verb form to which semantic class it belongs. In other words,
the distribution of each verb form must be semantically annotated, which provides
F(S), the frequency of each verb sense in the corpus, from which the frequency of
the verb sense in other constructions can be calculated: F(S ∧ ¬C)â•›=â•›F(S)â•›−â•›F(S ∧ C).
The last required frequency, F(¬S ∧ ¬C), is the frequency of other senses in
other constructions. In theory, it can be obtained by computing F(¬C)â•›−â•›F(S ∧ ¬C),
which thus comes down to finding F(¬C): the total frequency of other construc-
tions. In collexeme analyses of verbal constructions, it is often calculated from the
frequency of all constructions containing a verb, and thus equated to the frequency
of all verb tokens in the corpus minus the frequency of the construction (cf. Gries
et al., 2005; Stefanowitsch & Gries, 2003). However, I would argue that this strat-
egy is not appropriate in the case of verb-class-specific constructions. In theory,
F(¬C) corresponds to all cases in which another construction than the one under
study (C) is used in combination with a verb that could theoretically occur in C.
In practice, this set is hard to delimit (cf. Schmid, 2010, pp.â•›112–113; Schmid &
Kűchenhoff, 2013), but if one is dealing with a highly abstract construction such
124 Argument Structure in Usage-Based Construction Grammar
Verbs of ingestion refer to an event in which a living agent takes in some (normally
edible) substance and consumes it. When verbs of ingestion are combined with the
conative argument structure, the resulting clause usually conveys the idea that the
Chapter 5.╇ The importance of local generalizations 125
event occurs in a ‘bit-by-bit’ fashion, in that only a part of the substance referred
to by the at-phrase is ingested. Because the conative variant lacks the completion
that is conveyed by the transitive variant, the event of ingestion is open to repeti-
tion. Therefore the conative variant is more acceptable than the transitive variant
when the sentential context contains an element that indicates repetition, such as
the adverb again:
(2) a. He gulped at the beer again. (BNC GWF-1031)
b. #He gulped the beer again.
Van der Leek (1996) reports another possible semantic contribution of the conative
construction when it occurs with verbs of ingestion:
[U]sage of verbs of ingestion in the conative often seems to be motivated by a de-
sire to signal that no real attempt is (or even can be) made to carry out the action
to completion (p.â•›367)
Sentence (3) explicitly specifies the actual goal of the sipping (to be polite), and
thus entails that the agent Sandy has no real intention to consume the whole drink.
In (4), the adverb suspiciously marks that the agent seems to be distrustful of the
drink he has been served, as he indeed “doesn’t seem to like it”, and only drinks it
with some reluctance. Finally, in Example (5), the lack of intentionality of the agent
is explicitly mentioned through the adverb half-heartedly. It might be argued that
the transitive variant would not be crashingly bad in these examples, but the cona-
tive counterpart still sounds more felicitous, since the event described diverts from
a prototypical nibbling or sipping transitive event with regard to the intentionality
of the agent. Moreover, I did not find any transitive use of nibble or sip with an
adverbial indicating a lack of intentionality in the corpus.
5. Example taken from Van der Leek (1996, p.â•›367), originally from the Longman Dictionary
of Contemporary English.
126 Argument Structure in Usage-Based Construction Grammar
Events of cutting involve an agent moving a suitable instrument over the surface
of an object, and causing a rupture in the physical integrity of that object as a re-
sult. The semantic contribution of the conative argument structure with verbs of
cutting falls into the ‘lack of affectedness’ or ‘lack of completion’ semantic effects
discussed in the previous chapter, and often includes a ‘repetition’ component. As
we will see, this semantic contribution is reflected by the collexemes of the cutting-
conative construction.
In the event described by conative clauses containing a verb of cutting, the
contact made between the instrument and the referent of the at-phrase does not
bring about the effect that the transitive use of the verb would entail: the cutting
128 Argument Structure in Usage-Based Construction Grammar
either fails entirely, or is too minimal for one to consider that the object is indeed
cut. Hence, conative uses of verbs of cutting often convey the implicature that the
action is repeated. This is particularly evident in Example (6) below:
(6) We watched the men bundle up their parachutes and move off through the
dense undergrowth, chopping at it with jungle machetes. (BNC HR7-2964)
In this sentence, the agents attempt to make their way through vegetation by cut-
ting it down. In such a scenario, a single instance of chopping would not be enough
for this goal to be reached, as it would only produce a minimal effect on the tar-
get (given its nature). The chopping event thus consists of more than one blow.
Example (7) below presents a similar case with the verb saw.
(7) Killion felt her watching, and sawed clumsily at the bird. […] Eventually he
managed to hack off a wing, in two parts. The knife slipped and cut his finger.
(BNC HRA-4050)
This paragraph describes the cutting of some poultry, a task at which the agent
does not seem to be particularly skilled. If the transitive construction was used
in the first sentence, instead of the conative variant, it would entail that the cut-
ting task is complete. But the rest of the paragraph clearly indicates that it is not
the case, and that the cutting requires multiple attempts before some result is
eventually achieved.
Table 5.4 presents the collexemes of the cutting-conative construction. As
previously, significantly attracted or repelled collexemes are highlighted. The
analysis reveals four significantly attracted collexemes: hack, saw, chip and chisel.
As I will show, these five collexemes illustrate important aspects of the construc-
tional meaning.
Three of these collexemes, hack, saw and chisel, share a common semantic trait, in
that they are inherently repetitive: an event of hacking, sawing, or chiseling always
consists of several identical actions. Moreover, a single blow generally does not by
itself bring about the intended effect on the patient, e.g., cutting something to bits,
sawing a piece of some material apart, or chiseling wood into a given shape; the
movement must be repeated until the desired effect is obtained. Hence hack, saw and
chisel naturally support the semantic contribution of the cutting-conative construc-
tion in their conceptual semantics, i.e., both ‘lack of completion’ and ‘repetition’.
The next collexeme, chip, possesses comparable properties. In an event of chip-
ping, only a small piece of the patient is broken off, and chip does not in any case
support a truly holistic interpretation, i.e., an object that is chipped is only mini-
mally affected and keeps its overall physical integrity, compared to what happens
with true change of state verbs like break. Events of chipping must be repeated if
the patient is to be considered significantly affected.
The only significantly repelled collexeme in the list is cut. Its repulsion can be
explained by its status as a maximally neutral verb of cutting (and indeed the hy-
peronym of the whole class), which thus does not carry any semantic elaboration
that would promote its use in the conative construction. In addition, cut lends itself
to a holistic interpretation to a much greater extent than the attracted collexemes.
(8) Philip ran towards [the branch] and as he came round the slight bend in the
wide path he saw the boy tugging at it. He was trying to move it off the path.
‘I can’t shift it,’ he said to Philip. (BNC ABX-380)
In this example, the goal of the agent is to move the patient the branch from the
path that it obstructs, as made explicit by the subsequent sentence. This goal is not
reached, as the agent himself remarks. The transitive variant (he saw the boy tug-
ging it) would most probably be interpreted as meaning that the agent succeeded
in moving the branch, which would clash with the rest of the paragraph. Hence,
the conative variant is more felicitous.
Conative uses of verbs of pulling do not, however, necessarily entail that the
agent’s action has no effect. More generally, conative clauses encode the pulling
event but do not specify whether that event has an end state. Any effect that the
pulling might have may nonetheless be expressed in a subsequent clause, as in
Examples (9) to (11) below:
(9) The human pulled at the gate and swung it enough to allow itself inside.
(BNC HTH-1365)
(10) With his free hand he pulled at the cheap cloth of her blouse until it slipped
from her shoulders and bared her breasts. (BNC FU8-2050)
(11) Using both hands, Simon pulled at his leg. The mud released it, but without
his trainer. (BNC ACB3-258)
In all three examples, the two parts of the causal chain (cause and effect) are kept
separate and described by two different clauses: first the pulling event (the cause),
expressed by a conative clause, then the result of this action (the effect). If a transi-
tive clause was used to describe the pulling event, it would receive a cause-effect
interpretation, in that it would implicate an end state for the event. This end state is
provided by a conventional scenario determined by properties of the patient and/or
by the discourse setting, and in which some effect is brought about: the opening of
the gate in (9) and the removal of the patient from its current location in (10) and
(11). These implicatures would at best make the second clause seem redundant, if
not pragmatically odd. To prevent a cause-effect interpretation of the pulling event,
the conative variant is preferred over the transitive variant.
Another function of the conative construction in sentences (9) to (11) is to
insist on the difficulty of the task and on the efforts made by the agent to carry
it out. This is further supported in (9) by the fact that the effect on the patient is
lessened (as indicated by the enough-phrase), and in (11) by our commonsense
knowledge about objects getting caught in mud and the arduousness of taking
them out. The transitive variant would make this difficulty component less evi-
dent. The context also strongly suggests that the force is exerted on the patient for
Chapter 5.╇ The importance of local generalizations 131
an extended duration: the force is maintained until the relevant effect is brought
about; this is explicitly indicated in (10) by the until-clause. If a single pulling,
though extended, does not bring about the relevant effect, the conative clause also
allows an interpretation under which the pulling event is repeated. As the transi-
tive variant prototypically describes a punctual event, it lends itself less easily to
an interpretation of extended duration/repetition.
The property of conative clauses to cancel the default end state of a transitive
event of pulling is also useful when the pulling event is meant to have a different
effect from that found in the conventional scenario associated with the transitive
variant, as in Example (12) below:
(12) Wetting her fingertips, Vron plucked at the pages. (BNC H0M-2115)
The transitive counterpart of (12) (Vron plucked the pages) would be interpreted
as meaning that the agent removes pages off a book. However, the context and
particularly the adjunct wetting her fingertips make it clear that the goal of the
plucking is to turn the pages, not tear them apart. Using the conative construction
makes this interpretation all the more accessible, by preventing the default removal
reading of pluck. This explanation also applies to many other conative sentences
where the result sought by the agent is not strictly speaking an effect on the pa-
tient, or where no effect at all is intended (the ‘lack of intentionality’ component
mentioned in the previous chapter):
(13) Christina pulled at his arm to divert his attention. (BNC FRS-1245)
(14) She tucked her head down bitterly, her hands pulling anxiously at the lap of
her dress, then stood angrily. (BNC G04-1659)
(15) Evelyn was playing with a lace handkerchief as she talked, plucking at it ner-
vously. (BNC CN3-1052)
(16) She sat on the back door step, in the morning sunshine, bare feet on William’s
fawn and white belly, aimlessly pulling at his soft ears. (BNC HH9-3203)
In (13), the act of pulling is not meant to affect the pulled body part or its possessor.
As explicitly indicated by the purposive clause, the result intended by the agent
is to request the other protagonist’s attention. This goal does not correspond to
the default effect of the transitive variant (moving the arm or its possessor from
its current position), which motivates the use of the conative. In Examples (14) to
(16), neither a direct effect on the patient nor any other definite result is intended
by the agent: the latter acts respectively out of anxiety, nervousness, and idleness,
as indicated by the relevant adverbs. The transitive variant would be less felicitous
given this lack of intention on behalf of the agent.
132 Argument Structure in Usage-Based Construction Grammar
Table 5.5 lists the collexemes of the pulling-conative construction. The construc-
tion has three significantly attracted collexemes: tug, pick and pluck. The most
significantly attracted collexeme is, by far, tug. At first sight, it is not at all obvious
why this verb occurs so prominently in the conative argument structure, and con-
versely how it could provide an indication of the constructional meaning. The key,
I argue, lies in the typical properties of events of tugging, an information which can
be provided by lexicographic sources. Below is a list of definitions of the relevant
sense of tug taken from WordNet and several dictionaries:
(17) pull hard (WordNet)
to pull sth hard, often several times (OALD8)
to pull hard (Merriam-Webster)
To pull with great effort or force; to drag, haul. Often with at. (OED)
to pull with one or more short, quick pulls (Longman)
Finally, the occurrence of drag and pull among the repelled collexemes also lines
up with the meaning of the construction. Drag is more appropriately described as
a verb of accompanied motion (i.e., where both agent and theme move along the
same path, like bring) rather than a pure verb of pulling: it strongly presupposes
the motion of the patient, which makes it at odds with the conative construction.
Pull is of course the hyperonym of the semantic class; hence, it is arguably the most
neutral verb of pulling. Since it has no inherent semantic traits that particularly
favor the conative reading(s), its appearance as a repelled collexeme is expected.
6. Indeed, one of the two pulling senses of pluck in WordNet is in the same synset as the pulling
sense of pick, namely ‘pull lightly but sharply with a plucking motion’, e.g., he plucked the strings
of his mandolin. The other sense of pluck, ‘pull or pull out sharply’, e.g., pluck the flowers off the
bush, does not seem to widely differ from the sense synonymous with pick as shown by its gloss.
Pluck1, pluck2 and pick thus constitute a set of closely related verb senses.
134 Argument Structure in Usage-Based Construction Grammar
Verbs of hitting represent the largest of the four semantic classes considered in this
chapter. This class contains verbs that have either hit or strike as their hyperonym
in WordNet. Events of striking consist in an agent performing some movement
in the direction of a patient, aiming at forceful contact with it, usually with the
intention of affecting it in some way (causing it harm or damage).
One of the most commonly discussed semantic contributions of the cona-
tive construction with verbs of hitting is to prevent the entailment that the agent
makes contact with the patient. The event described by the conative variant is
construed as one in which the agent merely performs a particular motion towards
the patient, but whether that motion culminates in contact remains unspecified.
The conative variant differs in that respect from the transitive variant which en-
tails that contact is necessarily made. The ‘missed contact’ reading is exemplified
by (20) and (21) below:
(20) He asked, ‘What’s the problem?’ and in the same moment, struck at the man
stiff-armed, a karate blow. It should have come as a complete surprise, but the
man swayed, taking Ross’s fist on his shoulder. (BNC FS8-1809)
(21) Repulsed, he raised his arm to keep it from his eyes or mouth, and the flea
landed upon his hand. He slapped at it with his other hand, but it was beneath
his thumbnail before he could get to it. (BNC CRE-841)
In both cases, the context indicates that the agent fails to make contact with the
patient: namely, the subsequent clause or sentence specifies that the patient blocks
the agent’s attack. The transitive variant would conflict with this sentence, since
it would necessarily entail that contact is successful. It is important to stress that
the conative variant does not entail that contact fails: it just leaves this informa-
tion unspecified. In (22) and (23) below, the conative variant is used although the
next clause describes the effect of the blow on the patient, and thus presupposes
that contact was indeed made. These sentences are similar to Examples (9) to (11)
with verbs of pulling discussed in the previous section, in that the cause and effect
events are described in separate clauses.
(22) The Darkfall lightning scratched and flickered with its spider legs at the nearby
window… and Duvall kicked savagely at Barbara’s head, spinning her away
from Jimmy. (BNC G0E-3412)
(23) The man outside struck at the window again, smashing more of the wooden
frame, then he dropped the iron bar and snaked one hand inside, trying to
slip the catch. (BNC G0P-2800)
Chapter 5.╇ The importance of local generalizations 135
Preventing the entailment of contact is, however, not always the relevant interpre-
tation. Many instances of the hitting-conative construction occur in contexts where
contact with the patient is unlikely to fail. This is particularly obvious when the
patient is a large and stable landmark, such as a door in Example (24):
(24) Who on earth was thumping at the door like that early in the morning?
(BNC H8F-81)
In the context of these sentences, the agent is held captive inside a room from
which he attempts to escape, but fails to do so. The agent’s hitting the door is
clearly aimed at affecting it, by opening it, breaking it or bringing it down. The
transitive variant could implicate that this attempt is successful, which is why the
conative variant is preferred. In addition, the conative construction insists on the
difficulty of the task, by emphasizing the agent’s efforts or by suggesting that the
agent’s action is repeated.
Along the same lines, conative uses of verbs of hitting may differ from their
transitive variant in that they convey a lower degree of affectedness of the patient,
as with the verb pound in (27) below:
(27) Then, before anyone could stop her, Doreen rushed across the kitchen to slap
Lucy’s face and pound viciously at her head with closed fists.
(BNC HHB-4227)
In its transitive use, pound conveys a particularly strong force component, but this
force is attenuated in the conative variant. If the transitive variant had been used in
(27), it would entail far more damage on the patient than the conative variant does.
The conative variant in (27) conveys the idea of a pounding action with successful
contact, but downplays the effects of this action.
136 Argument Structure in Usage-Based Construction Grammar
The discourse context in both examples makes it clear that the agent’s action is not
aimed at hurting the patient (or rather, its possessor, since it is in both cases body
parts): in (28), the goal is to facilitate the penetration of an ointment into the skin,
and in (29), it is to prevent insects from biting (also, the agent’s action in (29) is
directed at herself, which makes an aggression scenario all the more unlikely). The
transitive variant would be less felicitous in these sentences because of the aggres-
sion scenario it would evoke.
This property of the hitting-conative construction is also relevant to sentences
which fall into the ‘lack of intention’ category discussed in the previous chapter, i.e.,
in which the agent does not act with the intention of bringing about any particular
result, as exemplified by (30) and (31):
(30) Luke swung away, punching at a box of tulips with one furious hand.
(BNC H8S-2974)
(31) Morthen hesitated, then struck at her brother’s shoulder with her clenched
fist. In turn he slapped her thigh and smiled at her […] (BNC HTM-943)
In (30), the agent acts out of anger; in (31), the agent acts playfully. In either case,
affectedness of the patient is irrelevant, and probably attenuated in (31). The lack
of intention of the agent would have been less obvious if the transitive variant had
been used instead of the conative variant.
To summarize, the semantic contribution of the hitting-conative construction
falls in either of the three following cases:
Table 5.6 lists the collexemes of the hitting-conative construction. It must be noted
that some of the verb sense frequencies in the table (third column) are not ‘real’
frequencies but mere approximations. The class of verbs of hitting is indeed too
large (12,847 instances in the corpus) for a manual annotation of the whole dataset
to be feasible. To reduce the amount of manual annotation, a random sample of 500
instances was taken for verbs that had more than 500 instances in the dataset. These
verbs are: bang (602), beat (1,372), hit (2,007), kick (1,186), pat (545), strike (1,990)
and tap (802). The frequencies displayed in the table were extrapolated from the
frequency of each verb sense in the random sample via a simple rule of three.
The significantly attracted collexemes are dab, lash, swipe, peck, hammer and
jab, and all of these verbs feature one or more particular semantic traits favored
by the construction. The verb dab, by far the strongest collexeme, is classified by
WordNet as a verb of hitting, though it is a very peculiar one. Contrary to more
typical members, dabbing involves little energy and is normally not aimed at af-
fecting the target, or at least not negatively. Rather, typical instances of dabbing
include using cloth to gather and remove a substance (like blood or tears), or
gently applying a substance on a surface (e.g., for medical or cosmetic purposes).
This typical lack of affectedness of the patient in an act of dabbing is in line with
the meaning of the hitting-conative construction: as I argued above, one of the
semantic effects associated with the construction is that the patient is not affected
by the agent’s action, or only minimally so.
The verb hammer originally refers to an act of hitting involving a hammer or
a similar tool as instrument; in that restricted use, typical things that can be ham-
mered include nails, metal sheets and other metallic goods. If anything, this use of
hammer typically entails repetition, i.e., just as with hack in Section 5.2.4, any event
of hammering normally involves multiple blows on the patient, since a single blow
does not suffice in affecting the patient in the intended way. For example, nails are
rarely properly hammered into a wall with a single blow, but rather inserted only
partly, and the hammering must be repeated as many times as necessary. Similarly,
a sheet of metal can never be shaped into any appropriate form with a single blow;
it has to be worked until the intended shape is arrived at. Of course, the verb in its
modern use is not restricted to exclusively describe acts of striking with a hammer,
but the ‘minimal effect’ and ‘repetition’ components found in the original meaning
of the verb arguably subsist, as modern dictionaries show, and the instrumental
component is echoed by the notion of a forceful and violent striking usually ac-
companied by loud noise (which many dictionaries gloss as ‘as if with a hammer’).
The remaining collexemes, lash, swipe, peck and jab, do not seem at first sight
to have much in common. However, their similarity becomes clearer when their
meaning is contrasted with that of other verbs in the class. As I briefly alluded to
at the beginning of this section, the various semantic contrasts among verbs of
138 Argument Structure in Usage-Based Construction Grammar
hitting can be shown to fall chiefly into two broad types. On the one hand, there
are impact-oriented verbs that specify the properties of the contact between agent
and patient, either in terms of intensity (e.g., pat vs. pound, respectively at the lower
and higher ends of the intensity scale) or some other property (for example, that
it is often accompanied by a loud noise, e.g., bang and thump). On the other hand,
there are agent-oriented verbs which do not particularly focus on the impact (or
only marginally so) but rather specify the kind of action performed by the agent:
they can describe a precisely defined shape of motion or put restrictions on the
kind of instrument or body part used by the agent (which also bears on the motion
performed). Lash, swipe, peck and jab all appear to be agent-oriented, as appears
from the definitions of the OALD. Lash (‘to hit a person or an animal with a whip,
rope, stick, etc.’) primarily restricts the kind of instrument used by the agent, which
Chapter 5.╇ The importance of local generalizations 139
in turn defines a particular shape of the agent’s blow. Swipe (‘to hit or try to hit sb/
sth with your hand or an object by swinging your arm’) is clearly identified with a
particular shape of motion in space: what makes an event of swiping is, above all, a
particular movement performed by the agent, namely a swinging blow (of the arm
or of an instrument). Peck (‘to move the beak forward quickly and hit or bite sth’)
refers to a particular part of the body of a bird and a movement performed with
that body part. The OALD does not record a hitting sense for jab, but the OED does
as ‘to give (a person) a stabbing blow with the fist’, which evidently refers both to
a definite shape of motion (‘stabbing’) and to a typical body part.
In sum, most collexemes of the hitting-conative construction are agent-oriented
verbs. As a matter of fact, there is a tendency for agent-oriented collexemes to clus-
ter towards the top of the list: other unattracted agent-oriented collexemes include
pummel (rank 7), swat (rank 8), scuff (rank 10), bat (rank 12), kick (rank 14) and
whip (rank 16). Agent-oriented verbs tend to occupy the top half of the list; the only
exception to that trend are punch (rank 21) and slap (rank 23). Impact-oriented verbs
chiefly occupy the bottom half of the list, for example beat (rank 24), bash (rank 18),
bang (rank 17) and whack (rank 13). It thus seems that agent-orientation is highly
correlated with the hitting-conative construction, which lines up with the focus that
the construction puts on describing what the agent is doing rather than the effects
that its action may have. The only impact-oriented verbs that are significantly at-
tracted to the construction, dab and hammer, are indicative of the constructional
meaning for different reasons: for dab, it is the lack of negative affectedness, and
for hammer, it is the necessary repetition due to the minimal effect of each impact.
Just like with the other verb-class-specific constructions, the maximally neu-
tral verb hit is, as expected, the most repelled collexeme of the hitting-conative
construction. Among the other repelled collexemes, beat and slap are also found:
the repulsion of these verbs can be explained by the fact that both verbs are fre-
quently associated with a punishment scenario which does not easily fit in with
the underspecification of contact or the lack of affectedness contributed by the
construction.
In this chapter, I presented findings pertaining to the relation between the mean-
ing of argument structure constructions and their verbal distribution. Following
up on the collexeme analysis of the conative construction presented in the pre-
vious chapter, which did not reveal a clear connection between constructional
meaning and verbal distribution, I suggested a different approach. Drawing on
evidence that lower-level constructions may be more relevant in a cognitively
140 Argument Structure in Usage-Based Construction Grammar
Hence, Stefanowitsch & Gries’ (2003) claims about the relation between the col-
lexemes attracted to a construction and the meaning of that construction are
clearly borne out for these four verb-class-specific instantiations of the cona-
tive construction. Namely, in each case the collexeme list clearly exemplifies the
principle of semantic compatibility and how this principle bears on usage. Verbs
with a meaning that lends itself particularly well to the interpretation sanctioned
by the construction are “attracted” by it: they are much more frequent in that
Chapter 5.╇ The importance of local generalizations 141
construction than chance alone would predict. Conversely, the repelled collex-
emes usually include verbs whose meaning is slightly at odds with the meaning of
the construction: eat for the ingestion-conative construction, cut for the cutting-
conative construction, drag and pull for the pulling-conative construction, and
slap, beat and hit for the hitting-conative construction.
In sum, collexeme analysis profiles the constructional meaning much better at
the level of each verb class than at the most general level. This does not mean that
collexeme analysis is ineffective for the conative construction taken as a whole: it is
just not particularly telling. The collexemes found for the overarching construction
are attracted because they are more compatible with the constructional meaning,
regardless of whether we define the latter at the most abstract level or at the more
specific level of verb classes. But the conative construction is so multifaceted when
taken at the most general level that it is much easier to understand why these verbs
are attracted collexemes and what this tells us about the meaning of the construc-
tion if we go down to the level of semantic verb classes.
On the theoretical side, these results shed some light on the nature of construc-
tional generalizations. As already mentioned in the opening section of this chapter,
a long-standing debate in constructional approaches to grammar is concerned with
which level of generalization best reflects speakers’ knowledge of constructions. If a
number of local generalizations account for what appears at first sight to be a single
general construction, this casts the question of whether an overarching construction
is needed at all, all the more so if the local generalizations provide a better account.
As it pertains to the usage basis of constructions, this is precisely the case with
the conative construction: to the extent that speakers attend to frequently occur-
ring verbs in some syntactic context, and use that information to “get a ‘fix’ on the
construction’s meaning” (Goldberg, 2006, p.â•›92), they can only exploit this lexical
semantic information at the level of verb-class-specific-constructions. Under this
view, a verb appears to be a collexeme of the general construction only because it
is, first and foremost, a collexeme of a verb-class-specific construction. In sum, the
findings presented in this chapter suggest that the conative construction is better
seen not as a unified whole, but as a cluster of low-level generalizations over similar
verb meanings (which might include other verb-class-specific constructions in ad-
dition to the four described in this chapter), in line with Croft’s (2003) proposal.
The psycholinguistic plausibility of at least some extent of higher-level gener-
alization (that is, at a higher level than verb classes) in the conative construction
should, however, not entirely be rejected. The fact that low-level generalizations
can determine the semantic contribution of the syntactic pattern for verbs of the
semantic class does not exclude the possibility of cross-generalizations between
different classes. First, if several distinct verb classes receive the same semantic
contribution (which is plausible, since the conative construction conveys a wide
142 Argument Structure in Usage-Based Construction Grammar
yet still limited range of meanings), they could form a single higher generalization,
which in turn could be used to produce new combinations. Second, patterns of
analogy between different classes might well play an important role in extending
the distribution of the construction (as argued by Bybee & Eddington, 2006) and in
helping speakers get at the correct interpretation, possibly forming generalizations
of intermediate scope. The generalizations accounting for the conative construc-
tion could well be centered on a few classes first, from which an abstract meaning
could be extracted and applied to other verbs and classes. Such a scenario is prob-
ably necessary to explain the inclusion of “orphans”, i.e., verbs whose semantic class
does not have any other representative in the distribution, such as shoot, which can
plausibly be analogized to a verb of hitting. Hence, if constructional generaliza-
tions start at a low degree of semantic generality, they may well provide the basis
for more abstract generalizations if similar form-meaning pairings are observed
among semantically different verb classes.
part iii
Alternations
chapter 6
In Part II, I have addressed how argument structure constructions come to be as-
sociated with meaning. As a result of this process, different constructions may also
share substantial aspects of their semantics, such that the use of verbs in certain
pairs of constructions involves relatively little variation in meaning, a phenomenon
commonly referred to as argument structure alternations. In this chapter, I suggest
that a thorough description of grammar should contain generalizations of com-
mon aspects of meaning shared by formally distinct constructions, such as those
involved in argument structure alternations.
The chapter is structured as follows. In Section 6.1, I define the concept of
argument structure alternations, and in Section 6.2, I discuss its status in construc-
tion grammar. It appears that very few construction grammarians consider alterna-
tions as grammatical units in their own right. I take issue with this position, and
in Section 6.3 I argue for a more explicit account of alternations in construction
grammar. Following a proposal by Cappelle (2006), I suggest that the variants of an
alternation can be described as allostructions, i.e., elaborations of a general event
schema that is shared by the two constructions and paired with an underspecified
form, capturing the fact that the grammar provides several ways to encode a given
category of events. In Section 6.4, I argue that allostructions can be used as sources
of productivity, i.e., they can generate new combinations of verbs and constructions.
As already noted in Chapter 2, most verbs occur in more than one syntactic envi-
ronment, like for instance the verb cook in the following examples:
(1) a. John cooked the rice.
b. The rice cooked.
c. John cooked us rice.
d. John cooked the rice sticky.
If one focuses on particular pairs of constructions, it often turns out that many
verbs may occur in both members of these pairs. For example, cook can be used
transitively as in (1a), and also in an intransitive construction as in (1b), where the
146 Argument Structure in Usage-Based Construction Grammar
subject argument corresponds to the direct object argument of the transitive use.
Many other verbs in English display the same grammatical behavior:
(2) a. John broke/shattered/cracked the glass.
b. The glass broke/shattered/cracked.
(3) a. John opened/closed/shut the door.
b. The door opened/closed/shut.
The two variants of the dative alternation clearly share much of their respective
meanings, namely that an agent initiates the transfer of an object or a message to
a recipient. In fact, as we will see in Section 6.3.2, the two are often interchange-
able, and the choice of one variant in a given situation has been shown to depend
on various properties of the alternating arguments themselves, such as their dis-
course accessibility (given vs. new), the animacy of their referents and their relative
lengths, rather than on semantic properties of the event itself; cf. inter alia Collins
(1995), Bresnan, Cueni, Nikitina, & Baayen (2007) and the references therein.
Chapter 6.╇ Alternations as units of linguistic knowledge 147
The study of argument structure alternations has received much attention in the
literature on argument realization, and admittedly there is much more work on
alternations to discuss than this chapter can reasonably accommodate (cf. Levin
& Rappaport Hovav, 2005, Chapter 7). A large body of research investigates the
factors that drive alternations and seeks to define the cognitive representations
and mechanisms that may give rise to them. Different linguistic theories give
different answers to this last question, which has much to do with their respec-
tive assumptions on the nature of grammar. In most constructional models of
argument structure, alternations as such have no independent theoretical status.
This is in striking contrast with earlier transformational (e.g., Emonds, 1972; Hall,
1965) or projectionist (e.g., Dowty, 1978; Jackendoff, 1975; Pinker, 1989) accounts,
in which specific operations were posited to derive one variant of an alternation
from the other variant. In the latter kind of approach, this position largely stems
from the assumption that only lexical items can convey meaning. Consequently,
any semantic difference between uses of a verb in various argument structures
1. Importantly, I am not concerned with how to deal with exceptions in this chapter. I assume
that both generalizations of argument structure and exceptions to these generalizations must
be part of a speaker’s grammar, but I will only focus on the former, leaving the latter for other
studies to account for. In particular, the process of statistical preemption has been suggested
in usage-based construction grammar to explain how learners acquire exceptions to argument
structure constructions (cf. Boyd & Goldberg, 2011; Goldberg, 2006, 2011; see also footnote 2
and Sections 6.2, 6.3.1, and 7.3.1).
148 Argument Structure in Usage-Based Construction Grammar
can only have a lexical origin, which is why projectionist approaches account for
multiple argument realization and its semantic aspects in terms of verbal poly-
semy. By the same token, since argument realization is driven by lexical semantics,
changes in argument structure can only originate in the lexicon, hence the idea
that argument structure productivity relies on the derivation of new lexical entries
(as mentioned in Chapter 2).
By contrast, constructional approaches importantly differ from projectionist
approaches in that they do not assume that the verb is the sole determinant of argu-
ment realization; rather, they take the view that certain aspects of argument struc-
ture must be attributed to the syntactic constructions themselves (cf. Chapter 2).
Construction grammarians also argue against derivations, whether they occur
at the syntactic level (transformations) or at the lexical level (lexical rules), and
instead put forward a monostratal view of grammar, in which linguistic forms of
any degree of schematicity are directly paired with meaning. A consequence of
this position is that alternation relations are seen with some suspicion and often
excluded from the inventory of grammatical units. Thus, in most constructional
models of argument structure, alternations as such have no independent theo-
retical status, and are rather seen as epiphenomenal. Hence, Goldberg (1995), in
dealing with the question of why some uses of the ditransitive construction can be
paraphrased with to (as per the dative alternation), reformulates the phenomenon
in constructional terms as follows:
The question that arises, on the account presented here, is not whether verbs are
allowed to undergo a lexical or syntactic rule that alters their semantic structure
or subcategorization frame, as it is typically taken to be. Rather, the question be-
comes: How are the semantics of the independent constructions related such that
the classes of verbs associated with one overlap with the classes of verbs associated
with another? (p.â•›89)
There are typically broader syntactic and semantic generalizations associated with
a surface form than exist between the same surface form and a distinct form that
it is hypothesized to be syntactically or semantically derived from. (p.â•›329)
Goldberg herself (2002) does not totally deny paraphrase relations any role in
grammar or language use. She acknowledges that their “statistical use […]
in actual discourse contexts is critical to unlocking Baker’s paradox of partial
productivity”,2 and that they “can also be seen to be relevant to on-line choices
2. Goldberg is referring here to the mechanism of statistical preemption, proposed to explain how
children learn that a verb cannot occur in some argument structure in view of its repeated use in
another less felicitous structure. See Goldberg (2006, 5.1) and Goldberg (2011) for more details.
150 Argument Structure in Usage-Based Construction Grammar
made in production” (Goldberg, 2002, p.â•›329). Thus, there does seem to be room
in construction grammar for generalizations over formally different constructions.
As a matter of fact, proposals to integrate alternations in constructional repre-
sentations are not unheard of. For example, Goldberg (1995, p.â•›91) posits a link
of “S-synonymy” between the variants of the dative alternation. Although little is
said about what such links truly are, how they emerge and what role they play in
grammar, language use and language development, they do capture the speakers’
awareness that two constructions have the same “descriptive” meaning, i.e., that
they can be used to describe the same set of situations.
A similar idea can be found in the treatment of argument structure alternations
in Frame Semantics (and, by extension, in FrameNet). As in construction gram-
mar research, alternations are not captured as such in Frame Semantics; rather,
they are considered as an epiphenomenon that results from the fact that the same
verb stem corresponds to several different lexical units (LUs), each evoking a dif-
ferent frame. For example, following Boas (2010, 2011a), the two variants of load
in the spray/load alternation (cf. Section 6.3.3) correspond to two distinct LUs: the
caused-motion variant (e.g., load hay onto the wagon) evokes the Placing frame,
while the with-variant (e.g., load the wagon with hay) evokes the Filling frame. The
observation that a verb enters into the spray/load alternation is thus translated in
Frame Semantics as a mere case of verbal polysemy (see also Nemoto, 2005 for
a similar conclusion). Yet, the relatedness of variants of an alternation may still
be captured through relations between frame. As Boas (2010, pp.â•›70–71) puts it:
While FrameNet provides no explicit link or connection between the valence pat-
terns of the two LUs, there exists a frame-to-frame relation between the frames
evoked by the two LUs, i.e., the Filling frame uses the Placing frame. Thus, syn-
tactic alternations are accounted for in terms of frame-to-frame relations and the
valencies of pairs of lexical units evoking frames that are semantically related.
The “uses” relation between the Filling frame and the Placing frame captures the fact
that “the endpoint of a filling event requires a number of placing events that tem-
porally precede this endpoint” (Boas, 2011a, p.â•›218). In sum, in Frame Semantics,
an argument structure alternation does not correspond to a relation between LUs
themselves, but may correspond to a relation between the frames they evoke,3 which
is only identified as such because many verb stems evoke both frames.
In conclusion, alternations are considered of little theoretical relevance in con-
structional approaches, which has much to do with the non-derivational character
3. See Iwata (2005, 2008) for a similar account of the locative alternation in terms of different
but related construals of the same abstract scene (cf. Section 6.3.3). García Velasco (2011) makes
similar proposals concerning the English causative alternation (e.g., He broke the glass vs. The
glass broke).
Chapter 6.╇ Alternations as units of linguistic knowledge 151
of the framework. This position is made clear in Goldberg’s (2002) surface gen-
eralizations hypothesis, according to which broad generalizations of form and
meaning (i.e., constructions) are of primary importance, and any alternations these
constructions are involved in should be disregarded. I take issue with this position
on the grounds that a grammar ignoring semantic and syntactic relations between
constructions certainly misses a large portion of what speakers know about their
language. In the next section, I will argue for an intermediate approach in which
both alternations and surface generalizations coexist in the grammar.
In line with the view outlined in the previous section, Cappelle (2006) argues for a
more explicit account of alternations in construction grammar that still preserves
the non-derivational character of the framework, drawing on the example of verb-
particle constructions in English, i.e., the two different orderings of transitive parti-
cle verbs, with the particle either directly following the verb (Verb-Particle-Object,
e.g., pick up the book), or placed after the direct object (Verb-Object-Particle,
e.g., pick the book up). Earlier transformational approaches do not fundamen-
tally distinguish the two different particle placements and relate them through a
transformation, whereas in constructional accounts, the two orderings are seen as
independent constructions which “exist due to their different functional motiva-
tions” (Gries, 2003, p.â•›140). While Cappelle agrees with Gries (2003, 2005) that
each verb-particle construction is associated with distinct discourse-functional
properties, he also contends that cutting the link between variants might be too
extreme an approach, since “just because the two orderings are not linked by a truly
Chomskyan transformation, it does not necessarily mean that language users are
not aware of their relatedness” (p.â•›12).
152 Argument Structure in Usage-Based Construction Grammar
Cappelle presents two arguments that this relatedness should be made explicit
in the grammar. First, there are a number of idiomatic verb-particle combinations
with a fixed object NP that can be used with either ordering (though one ordering
might be preferred); e.g., pull up one’s socks – pull one’s socks up. Since there is no
semantic difference between each ordering of these idioms (in the sense that both
receive the idiomatic meaning), the view that they should be stored as independent
and unrelated units lacks psychological plausibility (p.â•›13). The second argument
put forward by Cappelle relates to language acquisition: since there is apparently
no principled way to decide which orderings a given verb-particle combination
is compatible with, ruling out the unacceptable constructions constitutes a learn-
ability issue. As a possible solution, Cappelle mentions Goldberg’s (1995) idea that
speakers may collect indirect negative evidence through repeated exposure to the
use of a construction in contexts in which it is dispreferred (a mechanism referred
to as “statistical preemption” in her later work; cf. Goldberg, 2006, Section 5.1):
“the more often speakers do not use a given form which would be most suitable
in a particular discourse context, the stronger becomes the evidence for a child
that this form is not acceptable: the heard form pre-empts the expected form”
(Cappelle, 2006, p.â•›15). Cappelle notes that such a mechanism presupposes that
speakers notice the relatedness of constructions in the first place:4
For acquisition on the basis of this kind of indirect negative evidence to be pos-
sible at all, language learners should be able to see a semantic similarity between
the construction actually used and the construction they would have expected to
be used. This is rather obvious, but the importance of this necessary condition
cannot be overemphasized. Indeed, one cannot unlearn a certain unacceptable
construction simply by being frequently exposed to a construction that is not
closely linked to it. (p.â•›17)
Therefore, in order for this solution to the learning problem to be workable, rela-
tionships between constructions must be part of a speaker’s linguistic knowledge.
To account for such relations with the descriptive apparatus of construction gram-
mar, without recourse to transformational syntax, Cappelle suggests that the two
particle placement constructions should be viewed as linked via a “supercategory”
capturing what these constructions have in common, namely the fact that they
consist of a verb phrase containing a direct object NP and a particle, while leaving
the linear order of these constituents unspecified. This is represented in Figure 6.1.
4. As already mentioned, Goldberg (2002, p.â•›329) makes a similar observation about paraphrase
relations when she notes that “the statistical use of paraphrases in actual discourse contexts is
critical to unlocking Baker’s paradox of partial productivity.”
Chapter 6.╇ Alternations as units of linguistic knowledge 153
Figure 6.1╇ The transitive verb-particle construction with its two allostructions
(from Cappelle, 2006, p.â•›18).
The verb-particle constructions specify the linear order of the constituents in-
herited from the superconstruction; in Cappelle’s terminology, they are allostruc-
tions, i.e., “variant structural realizations of a construction that is left partially
underspecified” (p.â•›18), by analogy with the terms allophone and allomorph, which
also correspond to alternative realizations of a particular linguistic unit (namely, a
phoneme or a morpheme). Pushing the analogy further, Cappelle (p.c.) suggests
calling the generalization inherited by allostructions a constructeme.5 Importantly,
this kind of representation aptly captures both aspects of an alternation. The con-
structeme and the inheritance links to each allostruction capture the fact that
the constructions are similar and indicate at which level, and the allostructions
themselves may include further syntactic and semantic/pragmatic details as to how
they differ from one another, such as the kind of discourse-functional information
attributed by Gries (2003) to each particle placement construction.
Note that, in essence, the semantic relatedness of constructions could as well
be captured by direct relations between constructions, like Goldberg’s synonymy
links. However, one benefit of the allostructions model is that it does not require
to posit a new type of construction-to-construction relation, since it relies on the
more basic and widely accepted taxonomic relation. In addition, this model is more
flexible than synonymy links with regard to the description of how the construc-
tions are related: the constructemes capture the level at which constructions are
semantically equivalent and the allostructions specify exactly how these construc-
tions differ. In this respect, the semantic aspects of the allostruction model bear
some resemblance to the frame-semantic treatment of alternating verbs, which
are taken to correspond to two lexical units evoking distinct but related frames.
In a frame-semantic description, the explicit frame-to-frame relations capture in
5. Note that the application of the term ‘alternation’ to syntactic constructions was established
through the same analogy with phonological and morphological alternations between allo-
phones or allomorphs of the same phoneme or morpheme. In that regard, Cappelle’s termino-
logical suggestions make perfect sense.
154 Argument Structure in Usage-Based Construction Grammar
what respect the meaning of alternation variants are related. Beside the difference
in notation, this account diverges from the allostruction model in that while the
former only focuses on how the meaning of verbs occurring in different construc-
tions are related, the latter explicitly associates this difference with variation in the
form of constructions. Also, while both frame-to-frame relations and Goldberg’s
synonymy links can a priori relate only two frames or constructions, the number
of allostructions of a constructeme is not limited. That being said, some of the
frame-to-frame relations could be thought of as providing the basis for the seman-
tic generalizations that underlie allostructions.
In summary, the long-standing reluctance of construction grammarians to
include syntactic alternations in the set of entities stored in linguistic knowledge
probably stems from the fact that they were the hallmark of earlier transforma-
tional and derivational analyses, and therefore are seen with some suspicion by
proponents of a monostratal approach to grammar. However, a constructional ap-
proach should not necessarily exclude a description of the semantic similarity and
syntactic correspondences between constructions, as “regularities that speakers
can extract from a number of analogical usage events” (Cappelle, 2006, pp.â•›3–4).
Just like some construction grammarians acknowledge the existence of purely
formal generalizations (albeit controversially, cf. Goldberg, 2006, Chapter 8), i.e.,
generalizations of common formal features independently of a shared meaning,
like the subject-auxiliary inversion construction (Fillmore, 1999), I argue, in line
with Cappelle, that a thorough description of the “constructicon” should also in-
clude semantic generalizations that are (at least partly) independent of syntactic
form. Including alternations explicitly in the grammar serves to capture the speak-
ers’ awareness that some constructions are semantically similar and can be used
as alternative ways to encode a particular category of meanings, whether these
alternations are modeled as lexical rules (as in Pinker, 1989), as synonymy links
(as in Goldberg, 1995), or as allostructions (as in Cappelle, 2006). In the remainder
of this chapter, I adopt Cappelle’s allostructions model. In the next two sections,
I show how it can be applied to two well-known argument structure alternations:
the dative alternation and the locative alternation.
The variants of the dative alternation clearly share much of their respective meanings
and can largely be seen as paraphrases. Goldberg (1995, 3.4.2) analyzes the to-variant
as a metaphorical extension of the caused-motion construction, describing events in
which an agent initiates the change of location of a theme along some path. The mo-
tion component inherited from the caused motion construction is to be interpreted
through “a metaphor that involves understanding possession as the ‘possessed’ being
located next to the ‘possessor’, transferring an entity to a recipient as causing the
entity to move to that recipient, and transferring ownership away from a possessor
as taking that entity away from the possessor” (Goldberg, 1995, p.â•›89), motivated
by the fact that “giving prototypically correlates with movement from a possessor
to a recipient”. Through this metaphorical mapping, the so-called Transfer-Caused-
Motion construction acquires the same meaning as the ditransitive construction.6
However, because of the metaphorical inheritance relation between the to-
variant and the caused-motion construction, which motivates the use of a loca-
tive sentence pattern to describe events of caused possession, it could be argued
that the to-dative construction conveys a locative meaning in its own right, albeit
of an abstract nature, and is therefore not truly synonymous with the ditransitive
variant (cf. Pinker, 1989 for such an account). In other words, causing an object to
change possessors can be understood as causing it to undergo motion to a location
in an abstract space that represents relations of possessions, along similar lines to
Jackendoff ’s (1983) Localist Hypothesis. The initial motivation for associating the
to-variant with a caused-motion meaning is clearly the locative preposition to, which
suggests that the recipient is construed as the goal of a path. However, Rappaport
Hovav & Levin (2008) point out that there is only scarce evidence that the to-variant
actually makes reference to a path component, since the recipient argument of the
to-variant lacks many of the usual distributional properties of to-phrases describing
‘true’ goals. First, recipient arguments marked with to cannot be questioned by the
locative wh-word where; the phrase to whom must be used instead:7
6. The synonymy of the two variants of the dative alternation is also supported by distributional
facts: a collexeme analysis performed on the to-dative construction in the ICE-GB corpus reveals
that give is its most prominent collexeme. Under the assumption that prominent collexemes de-
termine the meaning of the construction (cf. Chapter 4), this means that the to-dative construc-
tion is also strongly associated with the ‘caused-possession’ meaning, just like the ditransitive
construction.
7. Examples (8) to (11) are taken from Rappaport Hovav & Levin (2008).
156 Argument Structure in Usage-Based Construction Grammar
The kind of path described by at and towards could arguably receive an abstract
interpretation in terms of possession. In the spatial domain, these unbounded
prepositions describe a trajectory that extends in the direction of the landmark
but does necessarily reach it;8 therefore, one could expect that they could be used
to describe an attempted (but not necessarily successful) change of possession. Yet,
these prepositions cannot replace to in the expression of events of giving. By the
same token, the to-phrase cannot be combined with other locative prepositional
phrases to describe a complex path, nor can it be modified by locative adverbials,
as shown by Examples (10) and (11) below:
(10) a. *Jill gave/offered the ball off the shelf to Bob.
b. *Jill gave/offered the ball from herself to Bob.
c. *Jill gave/offered the ball through the window to Bob.
(11) *Susan gave the ball all the way/halfway to Bill.
In sum, there is no independent motivation for assuming that the verbs entering
into the dative alternation should receive a different semantic representation for
each variant, which concurs with Goldberg’s analysis of the two constructions as
being essentially synonymous. Therefore, the variants of the dative alternation can
be characterized as allostructions: they correspond to different structural realiza-
tions of the same event-level meaning, in this case a transfer of possession. This is
illustrated in Figure 6.2.
X cause Y to have Z
NPX V { ?Y ?Z }
ditransitive to-dative
NPX V NPY NPZ NPX V NPZ to NPY
8. With the proviso that at only conveys this meaning in certain verbal constructions, as I
argued in Chapter 4.
Chapter 6.╇ Alternations as units of linguistic knowledge 157
While the dative alternation does not seem to be driven by purely semantic factors,
this does not mean that both variants are equally felicitous in any given context.
Most notably, the choice of one variant in a given situation has been argued to
depend on the discourse accessibility of the recipient and theme arguments. In
the double-object variant, the referent of the recipient argument is typically al-
ready given in the discourse context, while the theme argument tends to be new
information, and vice-versa in the to-dative variant (Collins, 1995; Erteschik-Shir,
1979; Thompson, 1990). This observation leads Goldberg (1995) to claim that the
ditransitive construction and the to-dative construction differ in their pragmatic
structure, in particular as to which argument receives focus: the recipient for the
to-dative construction, or the theme for the ditransitive construction. This account
explains why pronominal (and hence discourse-given) themes are usually ruled
out from the ditransitive construction,9 as in (12) below, and why the to-dative
construction is less felicitous with a pronominal recipient, as in (13).
(12) *John gave Mary it.
(13) a. ?Mary gave a brand-new house to him. (from Goldberg, 1995, p.â•›92)
b. Mary gave him a brand-new house. (idem)
the theme, and the theme argument tends to be discourse-new, lexical, indefinite
and longer than the theme. In the to-dative construction, the values of the proper-
ties of the recipient and theme arguments are exactly reversed: the theme tends
to be discourse-given, pronominal, definite and shorter than the recipient, and
the recipient tends to be discourse-new, lexical, indefinite and longer than the
recipient (see also Theijssen et al., 2013 for an exemplar-based modeling of the
two constructions along similar lines).
In sum, while the allostructions of the dative constructeme share a large part
of their constructional semantics, they appear to differ in terms of a number of
properties related to the recipient and theme arguments, constraining the range of
contexts in which each may be felicitously used. Bresnan & Ford (2010, p.â•›178) also
find that some verbs display strong constructional preferences for one or the other
variant that can win over the contextual factors determining the choice of variant
for other verbs; for instance, sell is biased towards the to-dative construction, and
teach is biased towards the ditransitive construction. These verb-specific biases
could be accounted for by positing different degrees of generalization for the dative
allostruction, in line with the emphasis that I put on lower levels of generalization
in Chapter 5. As Cappelle (2006) suggests for the particle placement alternation,
verb-specific dative allostructions could be defined in which the relative promi-
nence of contextual factors determining the choice of variant vary from verb to
verb, making the relative likelihood of either variant vary correspondingly. This
does not, however, preclude the possibility of a verb-general dative constructeme
generalized over verb-specific preferences.
The locative alternation11 offers two different ways of encoding an event of caused
change of location of some object, called the theme. In one variant (hereafter the
locative variant), the theme is realized as the direct object and the location is men-
tioned in a path expression (typically a prepositional phrase); cf. (14a) and (14a). In
the other variant (hereafter the with-variant), the location is realized as the direct
object and the theme as a prepositional phrase headed by with; cf. (14b) and (14b).
11. What I refer to here as the locative alternation is more often known in the literature as the
spray/load alternation, thus named with reference to the two verbs that are commonly used to
illustrate it. The spray/load alternation is one member of a family of several so-called locative
alternations (cf. Levin, 1993, 2.3) that offer different ways of encoding events of motion of a
theme vis-à-vis some location. For the sake of simplicity, I will use the term ‘locative alternation’
to refer exclusively to the spray/load alternation in this and the next chapter.
Chapter 6.╇ Alternations as units of linguistic knowledge 159
More so than the dative alternation, the locative alternation is often argued to in-
volve a tangible difference in meaning. The sentence pairs (14) and (15) exemplify
the so-called “holistic/partitive effect”, described by Anderson (1971, p.â•›389) as “a
matter of whether the whole of something is affected by the action described by
the sentence, or just a part of it is affected”. In other words, the with-variant with
spray and load seems to involve a holistic interpretation implying that the loca-
tion has been totally affected, whereas it need not be in the other variant. Hence,
(14b) implies that the wall is totally covered with paint, and (15b) implies that
the truck is full of hay.
The notion of affectedness is central to many accounts of the locative alterna-
tion which respectively characterize the locative variant as describing a change
of location, and the with-variant as describing a change of state (cf. Pinker, 1989;
Rappaport & Levin, 1988). Pinker (1989) paraphrases the locative variant as “X
causes Y to move into/onto Z” and the with-variant as “X causes Z to change state
by means of moving Y into/onto it”, with X standing for the agent, Y for the theme,
and Z for the location. In these descriptions, the meaning of the locative variant is
subsumed in the meaning of the with-variant, which in itself captures some degree
of semantic relatedness.
Goldberg (1995) provides a largely similar analysis of the variants of the alter-
nation in construction grammar. In her account, the locative variant instantiates
the caused-motion construction. The with-variant, however, is not really consid-
ered as a construction in its own right, since it is analyzed as an instance of the
transitive construction to which an adjunct with-phrase has been added. Thus,
since the transitive construction prototypically describes events of caused change
of state, Goldberg’s account is essentially based on the change of location vs. change
of state distinction found in previous analyses by Rappaport & Levin (1988) and
Pinker (1989). However, in this account, the with-adjunct does not present the
prepositional object as a theme, since the preposition with does not by itself convey
the idea of motion and is more likely to be taken in a near-instrumental reading in
the with-variant of the locative alternation. Hence, on that view, the with-variant
reduces to describing a change of state and does not necessarily make reference to
a change of location, contra Rappaport & Levin (1988) and Pinker (1989). Since, in
construction grammar, the constructional meaning determines the range of verbs
that may appear in a construction, the fact that the semantics of the verb includes
a change of location component should be irrelevant to the occurrence of that verb
160 Argument Structure in Usage-Based Construction Grammar
in the with-variant; only the fact that the verb can be construed as describing a
change of state is, as is the case for the transitive construction. This predicts that
all verbs occurring in the with-variant should allow the with-phrase to be omitted.
However, this prediction does not hold, as shown by the following examples taken
from Iwata (2008, pp.â•›14–15):
(16) a. Pat heaped her plate with mash-potatoes.
b. *Pat heaped her plate.
(17) a. He piled the shelf with the books.
b. *He piled the shelf.
(18) a. Pat crammed the jar with pennies.
b. *Pat crammed the jar.
Since the plain transitive construction is unacceptable with these verbs, as shown
in the (b) examples in (16) to (18), it follows that the same construction cannot
be instantiated in the corresponding with-variant examples, which are perfectly
acceptable. The (a) examples can therefore only be accounted for by assuming that
the with-variant corresponds to an independent construction, which I call the
with-applicative construction. The meaning of this construction does make refer-
ence to a theme argument, but not necessarily to a change of state. In fact, Iwata
(2008, p.â•›24) also notes that it is not always possible to characterize the patient
argument of with-applicative clauses as undergoing a change from an ‘old’ state
(e.g., ‘not cut’, ‘alive’, ‘not broken’) to a ‘new’ state (e.g., ‘cut’, ‘dead’, ‘broken’), as is
the case with typical change of state verbs like break, kill, and cut. Indeed, sentence
like He sprayed the wall with paint is compatible with a context in which it is hard
to imagine what change of state the spraying event might have brought about, e.g.,
if the patient argument the wall was already covered with paint prior to the event,
or is not fully covered after the event.
In sum, the entailment of a change of state (or affectedness), while relevant to
some cases, does not seem to be inherently associated with the with-applicative
construction, or only weakly so. Hence, the two variants of the locative alterna-
tion seem to be largely similar with respect to their constructional meaning. Yet,
I do not want to claim that they are synonymous in every respect. They do differ
semantically, but it appears that the difference lies at a more abstract level: while
they essentially describe the same verbal event, they reflect a different construal
of it. Both constructions refer to an event of caused change of location, but the
caused-motion construction construes the event as an action by the agent on the
theme, causing it to move, whereas the with-applicative construction construes it
as an action on the location, where the motion of the theme is merely the means
whereby this action is exerted. Note that this is a slightly different statement from
Chapter 6.╇ Alternations as units of linguistic knowledge 161
saying that the two variants differ in terms of whether the location argument is
specified to be actually affected. Defining the semantic difference in terms of con-
strual explains why there are pairs of alternating sentences that do not seem to
differ in their truth-conditional meaning. Either variant can potentially be used to
describe an event of caused-motion, as long as the construal imposed by the variant
does not clash with the meaning of the verb or with the specifics of the situation
at hand. In particular, the construal of the with-variant often calls for a significant
effect that the agent intends to bring about on the location. On this account, the
holistic interpretation can be seen as a mere pragmatic effect dependent on the
selection of the location as the entity acted upon.12
In sum, the variants of the locative alternation share a substantial part of their
constructional meaning, and can be therefore seen as two constructional options
for the description of events of caused motion involving alternate construals.
The caused-motion construction and the with-applicative construction are dia-
grammed in Figure 6.3.
12. Iwata (2008) offers a different account of the with-variant to explain why the implication
of a change of state does not always hold. Iwata maintains the earlier characterization of the
alternation in terms of the holistic effect by positing two verb-class-specific constructions (cf.
Croft, 2003 and Chapter 5 of this book), the cover-class-specific construction and the fill-class-
specific construction, which have the syntax of the with-variant and respectively convey that
the location argument has been totally covered or totally filled. Sentences that are semanti-
cally at odds with these constructions are claimed to instantiate different constructions. For
example, the sentence The vandal sprayed the sculpture with paint allows an interpretation in
which the statue is not fully covered with paint but its esthetic value is ruined even by the
slightest smudge; according to Iwata, this interpretation is sanctioned by the damage-class-
specific construction. In sum, while, in my account, whether and in what respect the location
might be affected is a mere pragmatic implicature that depends on world knowledge and
contextual information, in Iwata’s model it is part and parcel of the meaning of various con-
structions. I do not fully subscribe to this analysis, but it is far beyond the scope of the present
work to engage in a thorough discussion. I will therefore limit myself to the observation that
since Iwata’s account seems to ascribe any possible semantic effect that might be found in the
with-variant to constructional meaning alone (which necessitates more than one construc-
tion), it might have to contain an open-ended set of constructions, which not only seems
descriptively impracticable but also cognitively unrealistic. At any rate, Iwata’s account and
mine are not inherently incompatible: my characterization of the with-applicative construc-
tion as construing the event as an action on the location can be seen as an abstraction over
more specific and possibly pragmatically-derived constructional meanings like those posited
by Iwata. As argued in Chapter 5, this abstract construction can plausibly co-exist with the
lower-level constructions.
162 Argument Structure in Usage-Based Construction Grammar
NPX V { ?Y ?Z }
X cause Y to go Z
X cause Y to go Z X cause Y to go Z
X act on Y X act on Z
caused-motion with-applicative
In my study, I used the same experimental paradigm as Bencini & Goldberg with
a different set of sentences that includes the factor of alternations. Specifically, instead
of contrasting verbal vs. constructional sorting, I investigated whether the presence
of possible alternation relations has an influence on the way speakers categorize
sentences. Twenty-six native speakers of English were exposed to a set of sixteen sen-
tences instantiating three argument structure constructions: (a) the caused-motion
construction, (b) the ditransitive construction, and (c) the with-applicative construc-
tion. Importantly, the caused-motion sentences I constructed fall into either of two
types: (i) locative caused-motion sentences, describing the change of location of a
theme caused by an agent, and (ii) to-dative caused-motion sentences, which receive
an interpretation of transfer of possession through the metaphor ‘transfer of owner-
ship is physical transfer’ (cf. Section 6.3.2). Also, the to-dative sentences contained
verbs of throwing (throw, toss, chuck, kick), giving the to-dative sentences a double
reading. Since they describe physical motion, they can be categorized as instances of
the locative caused-motion construction, in which case the referent of the to-phrase
is a true goal. However, they also accept a caused possession reading and thus can be
categorized as instances of the more specific to-dative caused-motion construction,
in which the goal argument is also a recipient. Therefore, while the caused-motion
sentences of both types can be unified under a common meaning ‘physical transfer’,
each type is also semantically similar with one of the other constructions appearing
in the stimuli set. Namely, the to-dative type is similar to the ditransitive construc-
tion with respect to the shared notion of ‘transfer of possession’, and the locative
type is similar to the with-applicative construction in that they both describe events
of (physical) change of location. In fact, sentences of both types could be rephrased
with the other construction, with little variation in meaning.
As it turns out, subjects rarely relied on the caused-motion construction in
their sorting. They much more frequently presented a sorting solution in which the
variants of one of the alternations (either dative or locative) were together in one
group. In other words, they were overall more influenced by the semantic similarity
involved in either alternation than by the semantic similarity found between in-
stances of the caused-motion construction. In sum, alternation-based generaliza-
tions are reflected in the results of the sorting task much more often than purely
constructional ones. The group descriptions given by subjects in post-experiment
interviews largely confirm my interpretation of this quantitative trend, in that in
most cases the semantic explanation put forward by each subject to motivate their
sorting corresponds to the semantic commonality between variants of an alterna-
tion or instances of a construction.
These results indicate that speakers are able to formulate broader generaliza-
tions of a constructional meaning shared by formally distinct constructions. In
the experiment, they used the former more often than the latter in categorizing
Chapter 6.╇ Alternations as units of linguistic knowledge 165
sentences. However, returning to the central question of this section, it is not clear
whether these results support the claim that alternation-based generalizations are
actually stored in the mental grammar. The fact that speakers do perceive variants
of an alternation as closely related in meaning is certainly a prerequisite for the
actual storing of such cross-constructional generalizations, like the constructemes
of Figure 6.2 and 6.3, but it is in no way a sufficient condition. Indeed, it might be
argued that the generalizations that subjects made in the context of the experiment
were “ad hoc” categories (Barsalou, 1983) that result from conscious reasoning
about how the meaning of different sentences relate.
Such comments are in order, but it can still be argued that stored categories
should be more readily available than those created “on the fly”. Since alternation-
based generalizations were relied on much more often in the sorting task than
constructional ones, it is reasonable to hypothesize that they correspond to stored
generalizations. That being said, the question whether speakers posit such general-
izations outside of the context of a sorting task is admittedly in need of additional
evidence of other kinds, especially “on-line” evidence collected in tasks of language
comprehension or production. Fortunately, two recent priming experiments pre-
cisely provide this kind of evidence.
In cognitive psychology, priming refers to the phenomenon whereby prior ex-
posure to a stimulus influences response to another stimulus (either positively or
negatively) and is usually taken as evidence that the two kinds of stimuli are related
at some level in cognition. Priming effects have also been found in language. Notably,
there is substantial evidence for syntactic priming, whereby processing of a given
syntactic structure facilitates subsequent processing of the same structure. Priming is
considered by many psycholinguists as a valid method to tap into speakers’ linguistic
knowledge and to directly address questions pertaining to the cognitive organiza-
tion of grammar. More specifically, if the processing of a certain linguistic unit is
shown to influence the processing of another, this is taken as evidence that the two
must be related in some way in the mental grammar. Evidence for syntactic priming
has also been found in corpora, whereby the use of a particular syntactic structure
increases, other things being equal, the likelihood of the same structure being used
in subsequent discourse (cf. Gries, 2005, 2007; Szmrecsanyi, 2006).
Several recent studies report that syntactic constructions (in the construction
grammar sense of form-meaning pairs) can be primed (Chang, Bock, & Goldberg,
2003; Eddington & Ruiz de Mendoza Ibáñez, 2010). These results suggest that prim-
ing involves both form and meaning; however, these studies do not address whether
priming could occur on the basis of semantics alone, i.e., whether two formally dis-
tinct but semantically similar constructions could prime each other. Evidently, this
question directly relates to the view of alternations that I advocate in this chapter,
i.e., as generalizations over constructions fulfilling roughly equivalent functions.
166 Argument Structure in Usage-Based Construction Grammar
If there is any stored relation between such constructions, priming could offer a
way to reveal that relation. As a matter of fact, two recent priming studies provide
supporting evidence for the existence of such cross-constructional generalizations.
First, Goldwater, Tomlinson, Echols, & Love (2011) investigate priming effects
in the dative alternation with children. In their study, subjects were presented with
pictures depicting events that could be described using one of the constructions
involved in the dative alternation, and were asked to provide a linguistic descrip-
tion of each picture. For example, a picture showing two boys playing baseball
could be described by either The boy is throwing the catcher a baseball or The boy
is throwing a baseball to the catcher. The pictures were presented in either of three
conditions. In the two priming conditions, subjects were shown two pictures prior
to the task that were both described by the experimenter using either the ditran-
sitive construction or the to-dative construction. For example, the experimenter
would present a picture depicting a ‘show-and-tell’ school scene, and would say
The girl is telling her classmates a story in the ditransitive priming condition, or The
girl is telling a story to her classmates in the to-dative priming condition. After an-
other similar picture-description pair, the subject would be shown a third picture
and asked to describe it in one sentence. In the baseline condition, subjects had to
describe the pictures without prior priming. It is precisely this baseline condition
that enables Goldwater et al. (2011) to test for cross-constructional priming, by
comparing the production rate of each construction not only between priming
conditions, but also relative to the baseline.
As expected, Goldwater et al. (2011) find that their subjects tended to produce
more utterances matching the prime construction in either priming conditions
than in the baseline condition. But in addition, priming by either construction
also increased the likelihood of using the other construction compared to the
baseline condition with no priming. If there was only constructional priming, i.e.,
priming of both form and meaning, it would be expected that priming by a certain
construction would only increase the use of the same form-meaning pair. The fact
that semantically similar though formally distinct constructions prime each other
is evidence that, as Goldwater et al. (2011) conclude, “semantic representations
(independent of sequence) are primed in structural priming tasks” (p.â•›168).
In a similar study, Vasilyeva & Waterfall (2011) investigate the priming of
transitive active vs. passive constructions with speakers of Russian. They find that,
compared to the transitive prime condition, exposure to ‘canonical’ passives (i.e.,
containing a past participle and structurally parallel to the English passive) not
only increases the production of the same construction both by children and adult
speakers, but also the production of a number of other constructions, such as, for
example, active transitive constructions with the patient (in the accusative case) in
pre-verbal position (cf. Vasilyeva & Waterfall, 2011, pp.â•›5–6). Importantly, while
Chapter 6.╇ Alternations as units of linguistic knowledge 167
these alternative constructions are formally different from the passives primes, they
do fulfill the same discourse function, namely emphasizing the patient argument.
The authors conclude that “what gets primed is […] a particular way of looking at
and interpreting a given situation that is captured by the priming sentence” (p.â•›20):
in other words, a construal, independent of the form it is paired with.
The results of both studies are remarkably in line with the allostructions model.
They both show that formally distinct but semantically similar constructions can
prime each other, which suggests that they are indeed related at some level of
representation in the mental grammar of speakers. These results show that the
linguistic knowledge of speakers might well contain a higher level of generaliza-
tions composed of highly abstract constructional meanings detached from any
particular form, which is tantamount to the notion of a constructeme.
In the previous section of this chapter, I argued that argument structure alter-
nations deserve more consideration in constructional approaches to grammar.
I pointed out that alternations are not necessarily incompatible with the basic tenets
of construction grammar, since they can be accounted for in a non-derivational
manner by positing constructemes, i.e., generalizations of a common meaning
and an underspecified form over formally distinct constructions. I also presented
empirical evidence suggesting that speakers indeed store such generalizations. Yet,
the question arises as to why they would do so. After all, I claimed in Chapter 5
that grammatical knowledge is better seen as organized around generalizations of
limited scope rather than highly abstract schemas, a claim which is also substan-
tiated by a large body of research (cf. inter alia Boas, 2003; Bybee & Eddington,
2006; Zeschel, 2009). Since constructemes correspond to an even higher level of
abstraction than constructions, they seem at odds with this conception of gram-
mar. The question, therefore, that must be addressed, is what benefits these higher
generalizations would provide in accounting for the linguistic behavior of speakers,
over a grammar composed exclusively of individual constructions.
I suggest that speakers plausibly form cross-constructional categories for the
same reason as they form any category: because they are useful to them. In particular,
organizing the constructicon into groups of semantically related constructions pro-
vides straightforward pathways to productivity: they provide speakers with an indica-
tion as to what the possible forms of their language might be, in that the occurrence of
some verb in a particular allostruction triggers the expectation that this verb can also
be used in the other allostruction(s) of the constructeme. In this section, I elaborate
on this idea of alternation-based productivity in a constructional perspective.
168 Argument Structure in Usage-Based Construction Grammar
if its meaning involves a similar system of relations to that contained in the con-
structional meaning. Since the meaning of an argument structure construction
is extracted from relational similarities across instances (mostly stemming from
the verb), this is tantamount to saying that a novel combination of a verb with a
construction is formed on the basis of analogy with previously observed instances
of that construction (cf. Boas, 2003, 7.3 for a similar idea). Under that view, the
stored meaning of the construction defines the relational similarity that the novel
instances must bear to the previous instances. By way of example, let us consider
the following overgeneralization error:
(19) Don’t say me that. (Gropen et al., 1989)
the same morphological base. On the other hand, the absence of constructions and
therefore of construction-based productivity in Pinker’s model fails to account for
many productive uses that cannot rely on paradigmatic analogy. Marcotte (2005,
2006), in accounting for children’s overgeneralization errors in the causative al-
ternation, brings together insights from both Pinker’s (1989) model and more
recent constructional approaches, and posits a mechanism for argument structure
productivity that is similar to the present proposal. Marcotte assumes that children
not only record occurrences of particular verbs in particular constructions, but
also tally them into what he terms “construction paradigms”, i.e., “a tally of the
constructions a verb has been observed with” (Marcotte, 2006, p.â•›218). In other
words, children would record the information that sets of verbs enter into a par-
ticular alternation, and use this information to extend the alternation to other
verbs observed in one of its variants:
In effect, children would notice that verbs with certain lexical semantic properties
share significant parts of their construction paradigms, and reason that other verbs
with those lexical semantic properties therefore also share those parts. Children
engage in a process of analogical paradigm completion.
(Marcotte, 2006, p.â•›219, emphasis in the original)
As is clear from this quote, Marcotte’s model takes lexical semantic properties of
verbs into account in the analogical mapping that leads children to hypothesize
new uses of a verb. In the allostructions model, this semantic component is already
captured in the meaning abstracted into the constructeme.
Obviously, alternation-based productivity is hard to tease apart from construc-
tion-based productivity: since allostructions share (by definition) a substantial part
of their respective constructional meaning, any verb occurring in one allostruction
by virtue of semantic compatibility will also be to some extent compatible with the
other allostruction, and thus allowed to occur in that allostruction, along the lines
of construction-based productivity. Therefore, it might not be possible to deter-
mine a priori whether a productive use of a verb in a variant of an alternation is
an instance of construction-based productivity or alternation-based productivity.
More probably, both types of productivity could be involved in such cases, both
conspiring to motivate the productive use of the verb. If such is the case, productiv-
ity with variants of an alternation should be qualitatively different from productiv-
ity with constructions that are not involved in an alternation. As a matter of fact,
there is experimental evidence supporting this expectation. For instance, Brooks
& Tomasello (1999) taught the meaning of a nonce causative verb to children. The
verb was uttered in ten model sentences with either a get-passive construction (e.g.,
The car got meeked by Big Bird) or a transitive construction (e.g., Big Bird is meeking
the car). The experimenter then prompted the child to use that verb in a sentence
Chapter 6.╇ Alternations as units of linguistic knowledge 171
by asking her a question which was either neutral (e.g., What happened?) or which
created a discourse context which would bias the subjects towards answering with
an active transitive (e.g., What did Big Bird do?) or with a passive (e.g., What hap-
pened to the car?), in order to elicit productive use of the novel verb. Brooks &
Tomasello (1999) looked at whether the novel verbs were used in the construction
with which they were not previously witnessed (i.e., passive in the active transitive
training condition, or active transitive in the passive training condition). One of
their main finding is that the extent to which children generalized the verb to the
other construction did vary according to the training condition, in that they were
more likely to use a verb in the active transitive when it was previously witnessed
in the passive construction than the other way around.
In a similar study, Conwell & Demuth (2007) investigated three-year olds’
knowledge of the dative alternation. Their study reports that there also seems to be
a productivity asymmetry within the dative alternation. As in Brooks & Tomasello’s
(1999) experiment, Conwell & Demuth’s study made use of nonce verbs taught to
the children in the experimental session. Two novel actions were acted out with
toys before the child (throw an object with a catapult and transfer something with a
conveyor belt), who was encouraged to reproduce the action. Each novel action was
then labeled with a nonce verb (pilk or moop), which was first used to describe the
action to the child without any particular syntactic construction (There is a word for
what you just did; it’s ‘pilk’! Can you say ‘pilk’?). After the child repeated it, the verb
was used in one of the variants of the dative alternation (e.g., You pilked the cup to
Toby!). The children were then asked to re-enact the action with different toys and to
say what they were doing, thus eliciting use of the novel verb. Children’s responses
were coded as to whether the verbs were used either in (i) the same construction
as that they were witnessed with, or (ii) the other variant of the dative alternation.
Conwell & Demuth (2007) conducted two such experiments. In their first ex-
periment, both verbs were used in the same construction, such that half of the sub-
jects heard them in the double-object variant, and the other half in the to-variant.
As it turns out, almost all children in this experiment were largely conservative, in
that all sixteen of them but one used the novel verbs in exactly the same structure as
the one they were heard in. In the second experiment, Conwell & Demuth tried to
elicit a more productive behavior by relying on syntactic priming. In each trial, the
first verb was presented in one variant of the alternation, which therefore served as
prime, and the second verb in the other variant. The order in which variants and
verb forms were presented was balanced across subjects. This strategy was suc-
cessful, as can be seen in Figure 6.4 taken from Conwell & Demuth (2007, p.â•›170):
subjects did sometimes use the novel verb in the other variant. However, the extent
of productivity varies greatly according to the model construction. When the verb
was presented in the double-object variant, children used a novel prepositional
172 Argument Structure in Usage-Based Construction Grammar
construction for that verb roughly 50% of the time, whereas they used an innova-
tive ditransitive construction only 10% of the time when they first heard the verb
in a prepositional construction. In other words, subjects were far less likely to use
the verb productively when the model construction was the to-variant than when
it was the double-object variant.
0.8
% of utterances
0.6
Innovative
Model-like
0.4
0.2
0
Double object Prepositional
Model form
The asymmetry found by Conwell and Demuth in the dative alternation is not
straightforwardly predicted by purely construction-based productivity, since the
construction in which the verb was presented, and not only the meaning of the
verb, seems to have an impact on productivity. Under the assumption that the two
constructions essentially convey the same meaning, which is substantiated by the
discussion of the dative alternation in Section 6.3.2 and the results of the sorting
task reported on in Section 6.3.4, the novel verb should be equally compatible
with both constructions. Hence, there is no reason why speakers would be using
the same verb meaning strikingly more conservatively if they hear it in one con-
struction than if they hear it with the other construction. In sum, these findings
suggest that the productivity of constructions related by an alternation is indeed
qualitatively different from the productivity of individual constructions, which
gives credence to the idea of alternation-based productivity.
Finally, further evidence for the limits of a purely construction-based view of
productivity is provided by an artificial language learning experiment conducted by
Wonnacott, Newport, & Tanenhaus (2008). In their study, they created an artificial
language that subjects had to learn on the basis of an input whose distributional
properties were carefully controlled and manipulated across several experiments.
Their made-up language had a rudimentary grammar consisting of two transi-
tive constructions, with equivalent semantics but different syntactic positions for
the agent and patient argument of the verbs. Wonnacott et al. conducted several
Chapter 6.╇ Alternations as units of linguistic knowledge 173
6.5â•… Conclusion
In this chapter, I was concerned with generalizations in the grammar of verbs that
go beyond constructions. Stemming from the observation that many verbs can
be used in both members of particular pairs of constructions, argument structure
alternations are defined as alternative realizations of the same set of arguments of a
verb, usually accompanied by systematic variations in meaning. I noted that due to
their non-derivational character, constructional approaches do not include alter-
nations in the grammar. This position mainly stems from a dedication to describe
grammar in terms of generalizations of a common meaning and form, in line with
Goldberg’s (2002) surface generalizations hypothesis, thus disregarding generaliza-
tions of a similar meaning over distinct forms. I took issue with such an extremely
constructionist approach, by pointing out that a grammar ignoring semantic and
syntactic relations between constructions might not be a complete account of the
syntactic knowledge of speakers. I followed a proposal by Cappelle (2006) aiming
at integrating alternations into construction grammar by considering their vari-
ants as allostructions, i.e., elaborations of a formally underspecified higher-level
construction, called the constructeme, associated with the meaning shared by the
174 Argument Structure in Usage-Based Construction Grammar
to the other construction(s) with which this item was previously attested? Some
experimental findings suggest that it might be the case. As already mentioned
in Section 6.4, Conwell & Demuth (2007) report a noticeable asymmetry in the
productivity of the variants of the dative alternation: the children tested in their
experiment were more likely to display productive behavior with a novel verb when
this verb was presented to them in the ditransitive variant of the alternation than
in the to-dative variant. However, it is not entirely clear whether the critical factor
is indeed the model variant, as there might be alternative explanations as to why
one variant appears to be more productive than the other.
Conwell & Demuth (2007) discuss several possible explanations for their re-
sults. They first consider discourse- and length-related factors; as I mentioned in
Section 6.3.2, a large body of research indicates that the choice between variants of
the dative alternation can be largely explained by a tendency for speakers to men-
tion discourse-given referents first and discourse-new material next, and to put the
longest constituent at the end of the clause (among other parameters). Here, the
length factor is unlikely to have played a role, since both arguments were of the same
length in all cases. As to discourse accessibility, there was apparently no difference
in discourse status between the theme and recipient arguments in the experiment,
since neither had been previously mentioned. The only possible factor that could
have biased children towards one construction is when they chose to refer to one of
the arguments with a pronoun. However, since Conwell & Demuth (2007, p.â•›175)
generally observe that “children […] failed to use pronouns with any consistency”,
they could not conclusively determine whether pronominality did play a role.
Secondly, they consider a frequency explanation, whereby children could be
biased towards the to-dative because they hear it more often and are therefore more
familiar with it. This explanation is also ruled out, since the ditransitive construc-
tion is actually more frequent than the to-dative construction in child-directed
speech, as shown by previous corpus studies. The same studies report that the
acquisition of the ditransitive variant occurs earlier than or at the same time as
that of the to-dative. However, Conwell & Demuth (2007, p.â•›176) also note that the
less frequent to-dative is instantiated by a higher number of low-frequency verbs,
and in particular that long-distance transfer verbs (like throw) do not commonly
occur in the double-object form, which could possibly be the source of children’s
reluctance to use the semantically similar novel verbs in that variant.
Finally, Conwell and Demuth suggest that children may be biased towards a
construal of the recipients as true goals (i.e., locations), since these recipients were
toys and thus inanimate. Added to the locative nature of the novel verb meanings,
such a construal should indeed promote the use of the directional preposition to.
While this explanation makes intuitive sense, it cannot be validated without further
investigation, which calls for a replication of the experiment with verb meanings
that do not involve physical motion.
Chapter 7.╇ The usage basis of alternation-based productivity 177
The first goal of my experiment is to see if Conwell & Demuth’s (2007) results
can be replicated in different experimental conditions, and in particular to check
whether the asymmetry in the dative alternation is a fact about the grammar of
English in general, or just a tendency displayed by children at a particular devel-
opmental stage. For this reason, I recruited adult participants.
Secondly, I also wish to test whether the asymmetry is dependent on the mean-
ing of the verb. Two of Conwell & Demuth’s most convincing explanations revolve
around the fact that the novel verbs were verbs of transfer from a distance by physi-
cal means, and as such are suspected to be biased towards the to-dative variant
for semantic and/or distributional reasons. In my experiment, I investigate pro-
ductivity in the dative alternation with two semantically different kinds of verbs:
first, verbs of physical transfer similar to those used by Conwell and Demuth, and
second, verbs of communication.
Finally, I am interested in testing whether other alternations display a similar
asymmetry in productivity as the dative alternation. A straightforward candidate
is the locative alternation, also discussed in the previous chapter and shown to be
describable in terms of a system of allostructions, just like the dative alternation.
The locative alternation is also tested with two different kinds of verbs: change of
state verbs vs. change of location verbs.
Since the experiment aims to elicit productive use of novel verbs from adults, it
cannot rely on the same experimental procedure as Conwell & Demuth (2007),
which was intended to mimic adult-child interaction. However, there is another
way in which adult speakers (as well as children, albeit probably to a lesser extent)
can be taught the meaning of previously unseen words. Namely, they are able to
infer the meaning of a novel word from the context in which it is used. For ex-
ample, upon hearing John ate a gorp, any speaker can draw the conclusion that the
unknown word gorp most likely refers to an item of food. Such a context-based
inferential learning is arguably how speakers acquire most of their vocabulary,
especially less frequent and/or more semantically complex words.
In this experiment, participants were exposed to novel verbs that were used
in short stories. I composed each story myself, including in them contextual cues
178 Argument Structure in Usage-Based Construction Grammar
that strongly hinted at the intended meaning of the verb. All stories consisted of
four sentences and had the same structure. The first sentence provided basic back-
ground information, the second sentence (the critical sentence) described a trigger
event and contained the new verb used in one variant of an alternation, the third
sentence continued the story and was usually the one providing the most contex-
tual cues to the intended meaning of the new verb, and finally the fourth sentence
described a concluding event which was expressed using the other variant of the
alternation, so as to create a priming effect and promote productive uses of the
novel verb, as in Conwell & Demuth’s (2007) study. The task was presented to the
subjects as a language game where the aim was for them to guess the meaning of
made-up words. They were told that they had to read short stories, each containing
a novel word, presented on a computer screen one sentence at a time.
After they had read each story, the subjects had to fulfill three tasks. In the
verb recall task, they had to type the novel verb on the computer keyboard. In the
meaning decision task, they had to decide on the meaning of the verb by picking
the definition that seemed the most appropriate to them, among three possible
choices. Finally, in the sentence completion task, they had to answer a basic ques-
tion of the form What did X do? by completing a sentence prompt containing
the new word. For example, if the novel verb pell was presented to the subjects in
the sentence Sam pelled a very heavy box to Ted, they would have to answer the
question What did Sam do? by typing a continuation of the sentence prompt Sam
pelled…. Crucially, the sentence prompt could be continued either by using the
model construction (here, the to-dative), which amounts to a conservative use, or
by using the other variant of the relevant alternation (here, the ditransitive), which
constitutes a productive use. The dependent variable in this experiment is the rela-
tive proportion of conservative vs. productive uses of novel verbs, according to the
model construction and the meaning of the verb.
While the first two tasks were mostly intended to make the subjects call the
new verb to mind and reflect about its possible meaning, the third task was in-
tended to make them actually use this verb in a sentence, thus possibly eliciting
productive uses. I tried to avoid that subjects produce rote repetition of the original
sentence by instructing them explicitly that the point of the task was not for them
to memorize the sentence in its exact formulation, but rather to produce a sentence
describing what happened in the story which could use an alternative wording as
long as it made sense to them and answered the question appropriately. In that
connection, a secondary role of the first two questions was to help minimize a per-
fect recollection of the critical sentence by providing some intervening distraction
between the end of the story and the third question.
The full procedure for each trial is summarized in Figure 7.1, exemplified with
the novel verb pell, intended to mean ‘transfer with a conveyor belt or a similar
Chapter 7.╇ The usage basis of alternation-based productivity 179
device’. This procedure was implemented in the E-Prime 2.0 experiment develop-
ment environment. After signing a consent form and filling out a questionnaire,
subjects were shown to a computer where the whole experiment took place. They
received the following instructions displayed on the computer screen:
In this experiment, we will play a language game. For this game, we have invented
new words. The aim of the game is for you to guess what these words mean.
We have used these words in short stories which will be presented to you one
sentence at a time. Each sentence will appear for a limited time. Read them care-
fully and try to spot the new word and guess what it means.
We will then ask you three questions: (1) what was the new word?, (2) what does it
mean? (you will have to choose among three suggested definitions), (3) a question
about how the word is used in the story.
First sentence Ted and Sam were testing the new machines.
Critical sentence
Sam pelled a very heavy box to Ted.
(here, pell + to-dative)
Short story
When the conveyor belt stopped,
Third sentence
Ted removed the box.
Priming sentence
Sam wrote their boss a positive report.
(here, ditransitive)
What was the new word in the short story you have
Verb recall task
just read?
They were then presented with two practice items for them to get used to the procedure.
These practice stories had the same structure and tasks as the actual items of the experi-
ment, but none of them contained sentences with a variant of the dative or locative
alternation. After they completed the practice items, the experiment proper started.
Each participant was presented with the same sixteen short stories. The stories
were presented in semi-randomized order for each subject, in that I prevented that
two stories featuring the same alternation (dative or locative) followed each other
in the list. As already mentioned, the four sentences in each story were presented
one at a time on the computer screen. Each sentence appeared for a limited time,
which I wanted to be long enough for subjects to read it completely and retrieve
relevant contextual cues, but not too long so as to prevent them from being able
to memorize sentences by rote (especially the sentence containing the novel verb).
In preliminary tests, I determined that a simple heuristic of 500 msec per word
in the sentence seemed appropriate; as it turned out, most subjects reported that
they did not find this rate either too slow or too fast. The three tasks following each
story were given in the order and with the instructions specified in Figure 7.1.
The order in which the definitions were listed in the meaning decision task was
randomized in each trial. All responses to the three tasks were provided by means
of a computer keyboard and were recorded by the E-Prime program. The whole
experiment usually took about fifteen to twenty minutes.
7.1.3â•… Stimuli
The stimuli set is based on two alternations: the dative alternation and the loca-
tive alternation. For each alternation, two semantic classes of verbs were tested.
For the dative alternation, these classes are verbs of physical transfer, describing
the caused motion of a theme along some path to a recipient through physical
means (similarly to the novel actions used by Conwell and Demuth), and verbs of
communication, describing the transfer of a message to a recipient using a spe-
cific instrument. For the locative alternation, the two classes correspond to what
Pinker (1989) term content-oriented verbs versus container-oriented verbs. Verbs
in the former class primarily describe the manner in which the theme is moved
in or applied on the location, whereas verbs in the latter class primarily describe a
particular result that the agent intends to bring about on the location.
I created four novel verb meanings for each semantic class. In each stimuli list,
two novel verbs of each class were presented in one variant of the relevant alterna-
tion, and the other two in the other variant. The assignment of variants to verb
meanings was balanced across subjects by constructing two types of lists. In type
A lists, the same two verbs meanings in each class were used with one variant, and
the other two with the other variant. In type B lists, this assignment was reversed.
Chapter 7.╇ The usage basis of alternation-based productivity 181
The priming sentence of each story also varied according to the type of list, so that
the variant of the alternation it exemplified was different from the one in which
the novel verb was used in the critical sentence. Half of the subjects received type
A lists, the other half type B lists. By doing so, I collected the same amount of data
for each verb meaningâ•›×â•›construction combination.
Sixteen nonce verb forms in the past tense were used as novel verbs: greemed,
norped, pelled, mooped, tonked, pilked, keated, cronted, wugged, blicked, speffed, hel-
ted, tusted, preated, grashed, and frodded. The first seven were taken from research
articles (Gropen et al., 1989; Zwicky, 1971). For the rest of them, I generated mono-
syllabic base forms from the MCWord Orthographic Wordform Database available
online,1 and put these forms in the past tense. MCWord is based on the CELEX
lexical database of English; the web interface includes a generator that combines
graphemes extracted from CELEX to form non-words, following patterns found in
actual words. By using this generator, I ensured that the non-words comply with the
phonotactics of English, and therefore sound like possible English words. For each
participant, the verb forms were randomly assigned to the sixteen verb meanings.
The novel verb meanings of each class are described below. For reasons of
space, I do not include the full stories in this section, but they can be consulted
(along with the alternative definitions given to subjects in the meaning decision
task) in the Appendix.2
1. http://www.neuro.mcw.edu/mcword/
2. It has been brought to my attention (Maarten Lemmens, personal communication) that a few
sentences used in four of my stories (and one definition) were not formulated in fully idiomatic
English. As a matter of fact, the stories had indeed not been checked by a native speaker before
the experiment started. However, I do not believe that such mistakes could have significantly
influenced the performance of the subjects in the sentence production task, because the critical
sentences and the prime sentences were all error-free. At worst, unidiomatic formulations in
the other sentences might have hindered the identification of the intended sense of the verb, but
since in the analysis of the results of the sentence completion task I only kept the responses in
which the subject chose the definition that I actually intended, this problem is arguably incon-
sequential as far as measuring productivity is concerned.
182 Argument Structure in Usage-Based Construction Grammar
spreading of pollen among flowers: they collect pollen on their body in a flower
they gather from, and spread them to the other flowers they will visit. Meaning 2
refers to a kind of repeated rubbing, accompanied by the theme wearing off on the
location. Meaning 3 describes a way to apply a fluid by pressing and squeezing a
piece of cloth soaked with it. Meaning 4 refers to a particular way (left unspecified
in the sort story but taken to be of some artistic value) of moving an instrument
against a surface in order to apply a substance on it.
7.2â•… Results
Forty native speakers of English (20 male, 20 female) participated in the experi-
ment. They were aged between 20 and 35 (23.45 on average) and were all students
at the University of Freiburg. About half of them (24) came from the United States,
seven from the United Kingdom, three from Canada, three from Australia and three
from New Zealand. They were rewarded with a monetary compensation for their
participation, except three who took the experiment for course credit. As indicated
in the last section, the first twenty were given a type A list, and the other twenty
were given a type B list. From each of these subjects, 16 verb recalls, 16 meaning
decisions, and 16 sentence completions were collected, resulting in 640 datapoints
in each category. In this section, I report the results of the experiment by focusing on
two tasks: first, the meaning decision task, and then the sentence completion task.
ditransitive construction, 143 (89%) for verbs modeled in the to-dative construc-
tion, 115 (72%) for verbs modeled in the caused-motion construction, and 122
(76%) for verbs modeled in the with-applicative construction. Even though several
of the alternative definitions selected by participants are semantically similar to the
“correct” ones, and should therefore not induce dramatic differences in argument
realization, restricting the dataset in this way ensures that all the responses taken
into account were provided by subjects sharing the same semantic assumptions.
Table 7.5╇ Success rates in the meaning decision task for each novel verb meaning.
Intended verb meaning Success
Dative verbs of physical transfer
1. transfer from a distance using a conveying device 37 (92.5%)
2. hurl from a distance and over obstacles 33 (82.5%)
3. send through a closed conveyor 33 (82.5%)
4. send something by blowing through a cylindrical instrument 35 (87.5%)
Total 138 (86.25%)
Dative verbs of communication
1. tell something with the Internet 38 (95%)
2. tell something with a broadcasting system 37 (92.5%)
3. tell something by using chemicals 38 (95%)
4. tell something with an amplifying device 33 (82.5%)
Total 146 (91.25%)
Locative content-oriented verbs
1. apply a substance by flying near something 28 (70%)
2. apply a substance by rubbing something repeatedly 13 (32.5%)
3. apply a substance by pressing and squeezing something 31 (77.5%)
4. apply a substance following a specific movement or shape 38 (95%)
Total 110 (68.75%)
Locative container-oriented verbs
1. obstruct a passage by placing things appropriately 38 (95%)
2. arm a ranged weapon by placing a projectile 37 (92.5%)
3. make a surface opaque by applying something 18 (45%)
4. make a container to not stick by applying some substance 34 (85%)
Total 127 (79.38%)
The results of the sentence completion task were coded according to the construc-
tion with which the novel verb was used. If it was the same as the one the verb
was used with in the short story, the sentence was coded as “same variant”. If it
186 Argument Structure in Usage-Based Construction Grammar
was the corresponding variant of either the dative or the locative alternation, the
sentence was coded as “other variant”. If it was neither of these constructions, the
sentence was coded as “unrelated”. To be regarded as “same variant” or “other vari-
ant”, the sentence completion had to contain both postverbal arguments (theme
and recipient for the dative verbs, and theme and location for the locative verbs);
other uses were coded as “unrelated”, and were ignored in the following analysis.
This left 101 tokens in the ditransitive presentation condition, 109 in the to-dative
presentation condition, 93 in the caused-motion presentation condition, and 82 in
the with-applicative presentation condition. The distribution of subjects’ produc-
tions according to the coding criteria is summarized in Table 7.6.
The distribution of subjects’ productions of the novel dative verbs (verbs of physi-
cal transfer on one hand, and verbs of communication on the other hand) in the
two presentation conditions (ditransitive vs. to-dative) is plotted in Figure 7.2,
collapsed across all participants. The error bars give the standard error of each
distribution, in order to provide an indication of variability across subjects.
1.0 1.0
% of dative productions
% of dative productions
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
Model: Model: Model: Model:
ditransitive to-dative ditransitive to-dative
Verbs of physical transfer Verbs of communication
Figure 7.2╇ Distribution of dative productions for each verb class and
in each presentation condition.
Chapter 7.╇ The usage basis of alternation-based productivity 187
As can be seen from this diagram, in the to-dative condition (i.e., novel verb pre-
sented in the to-dative construction), subjects were largely conservative with both
verb classes: they most frequently used the novel verb in the same construction
that they have just witnessed it with, and only rarely used it in the ditransitive
construction (physical transfer: paired t(36)â•›= 6.4293, pâ•›<â•›0.0001; communication:
paired t(39)â•›=â•›4.1284, pâ•›= 0.0002). However, in the ditransitive condition, subjects
displayed a productive behavior with both classes: they used novel verbs of com-
munication about equally often in either variant (the difference in frequency is
not significant: paired t(39)â•›=â•›−1.1355; pâ•›=â•›0.2631), and they used novel verbs of
communication even more often in the to-dative construction than in the model
construction (paired t(36)â•›=â•›−5.6183; pâ•›<â•›0.0001).
It might be objected that what we observe is not necessarily true productiv-
ity, but could be a mere instance of constructional parallelism, since I purposely
employed syntactic priming, using the other variant of the alternation in the last
sentence of each short story, in order to promote the use of that variant in the
subjects’ productions. In other words, it can be argued that when subjects used
the other variant, they simply did so because it was the last construction they had
heard, and not necessarily because of any assumptions about the productivity of
this construction. It is certainly true that the subjects were somehow artificially
prompted to be productive. We might observe different quantitative trends if I had
not relied on syntactic priming, but in fact, there might not have been any produc-
tive behavior at all, as in Conwell & Demuth’s (2007) first experiment. Be that as it
may, if syntactic priming was the sole explanation of the subjects’ responses, to the
exclusion of syntactic productivity per se, this would not explain the asymmetry:
why would one construction (here, the to-dative) prime itself better than another
construction (the ditransitive)? In sum, there appears to be more to the behavior
of participants than the mere repetition of a previously heard construction.
The reported results are similar to those reported by Conwell & Demuth
(2007), which shows that the asymmetry in the dative alternation is not only found
in language acquisition but is also displayed by adult speakers. Besides, they show
that the asymmetry in the dative alternation is not restricted to verbs describing
a transfer by physical means, even though the extent of that effect seems to vary
according to the meaning of the verb. This rules out Conwell & Demuth’s (2007)
suggestion that the bias towards the to-dative might be due to the fact that the
children participating in their experiment could have construed the recipient in
the enacted action as a true goal in a physical sense. Even though this semantic
possibility displayed by verbs of physical transfer may arguably contribute to the
strength of the effect (subjects did favor the to-dative construction in both presen-
tation conditions when they had to use these verbs), it is not at all necessary for this
188 Argument Structure in Usage-Based Construction Grammar
effect to arise in the first place. In sum, the asymmetry in the dative alternation can-
not be fully explained by a possible bias stemming from the meaning of the verb.
The distribution of subjects’ productions of the novel locative verbs (content-
oriented verbs on the one hand, and container-oriented verbs on the other hand)
in the two presentation conditions (caused-motion vs. with-applicative) is plotted
in Figure 7.3, collapsed across all participants.
1.0 1.0
% of locative productions
% of locative productions
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
Model: Model: Model: Model:
caused-motion with-applicative caused-motion with-applicative
Content-oriented verbs Container-oriented verbs
Figure 7.3╇ Distribution of locative productions for each verb class and
in each presentation condition.
As it turns out, the locative alternation does not display at all the asymmetry found
in the dative alternation. In both presentation conditions, subjects tended to use the
novel verb in the construction with which it was presented to them, and only rarely
generalized it to the other variant. This behavior is observed with both classes of
verbs, which means that, contrary to what might be expected, neither the content-
oriented verbs nor the container-oriented verbs have a particular preference for the
caused-motion construction or the with-applicative construction respectively. The
frequency difference is statistically significant for container-oriented verbs in both
conditions (caused-motion: paired t(33)╛=╛2.7159; p╛=╛0.0104, with-�applicative:
paired t(33)â•›=â•›5.8495; pâ•›<â•›0.0001). However, it is only marginally significant for
content-oriented verbs in either condition (caused-motion: paired t(36)â•›=â•›1.6408;
pâ•›=â•›0.1095, with-applicative: paired t(36)â•›=â•›1.8241; pâ•›=â•›0.0765). It is not clear how
this difference can be explained, but it could simply be due to the fact that subjects
were less successful at identifying the ‘right’ definition of content-oriented verbs
than of container-oriented verbs (as I reported in Section 7.2.1). Hence, there are
fewer data points for the former than for the latter, and less statistical power to
detect significant differences.
Chapter 7.╇ The usage basis of alternation-based productivity 189
These examples demonstrate that productivity in the dative alternation can be mo-
tivated by the same factors that drive the selection of variant for alternating verbs.
In all of them, the recipient argument is referred to as a pronoun; additionally, in
(1b) and (2a), the theme argument is markedly longer than the recipient argument.
Although the to-dative variant would not be crashingly ungrammatical, the use of
the ditransitive variant instead is in line with Bresnan et al.’s (2007) finding that
in such a case, the latter variant is preferred for alternating verbs. In the sentence
completion task of my experiment, subjects were left free to use pronouns and ei-
ther definite or indefinite noun phrases, and there was no limit to the length of the
constituents they could use as arguments of the verb. The question must therefore
be asked whether such choices could have influenced their responses with respect
to the variant of the dative alternation that they used.
190 Argument Structure in Usage-Based Construction Grammar
The following examples taken from subjects’ productions are clear cases in
which the relative length of the arguments constrains the choice of one variant
over another. In the examples under (3) below, the recipient argument is longer
than the theme argument by many syllables (and also more syntactically complex),
which makes the to-dative variant more appropriate than the ditransitive variant:
(3) a. The allies tusted [food]TH to [their friends in the city]REC.
b. The allies cronted [supplies]TH to [people they were helping]REC.
c. Paul preated [a letter]TH to [Laura who was standing on the balcony]REC.
Since these examples could hardly be formulated with the other variant, it is not
clear whether they are “genuine” cases of productive or conservative uses. Examples
(4) and (5) illustrate another constraint: in these cases, the speaker chose to refer to
one of the arguments with a pronoun, which makes the other variant inappropriate,
or at least less likely. In the examples in (6), both the pronominality constraint and
the length constraint favor the use of the ditransitive variant.
(4) a. Tom pelled [it]TH to [Laura]REC.
b. Paul helted [it]TH to [the woman]REC.
c. Sam speffed [it]TH to [Tom]REC.
(5) a. Harry tusted [her]REC [the message]TH.
b. Tom speffed [her]REC [the plan]TH.
c. Harry cronted [her]REC [the news]TH.
d. Jack frodded [her]REC [a package]TH.
(6) a. Harry wugged [her]REC [a cheeky facebook message]TH.
b. Harry pelled [her]REC [what her grade was]TH.
Nevertheless, it seems unlikely that length and pronominality alone could account
for the productivity biases, since only sixteen examples (out of 210) were found in
which these factors could be thought responsible for choosing one variant over the
other. Animacy should also not have played a role, since in all cases the recipient
was animate, which, according to Bresnan et al.’s (2007) model, should favor the
ditransitive variant; yet an overall preference for the to-dative was found. Also,
since both arguments were already mentioned in the story, they should not widely
differ in terms of discourse accessibility. However, the influence of definiteness
is more difficult to gauge and tease apart from the aforementioned factors. Also,
while the examples listed under (3) above represent cases where it is clear that
length played a role in the selection of a dative variant, this role is more difficult
to evaluate when differences in length are less striking. All in all, the influence
of contextual factors and their cumulative effects should be considered more
systematically.
Chapter 7.╇ The usage basis of alternation-based productivity 191
In order to do so, I used Bresnan et al.’s (2007) data extracted from the
Switchboard corpus to fit a logistic regression model, following the procedure de-
scribed in Johnson (2008, 7.4),3 with the following predictors: meaning of the verb
(transfer, future transfer, prevention of possession, and communication), discourse
accessibility of the recipient (given vs. not given), discourse accessibility of the
theme, pronominality of the recipient (pronoun vs. non-pronoun), pronominal-
ity of the theme, definiteness of the recipient (definite vs. indefinite), definiteness
of the theme, animacy of the recipient (animate vs. inanimate), and difference
in length between recipient and theme (expressed as the difference between the
logarithm of the length in graphemic words of the recipient and the logarithm of
the length of the theme). I coded the 210 sentences produced by subjects for these
nine variables. Verbs of physical transfer were coded for class as ‘transfer’, and verbs
of communication as (obviously) ‘communication’. For all sentences, the discourse
accessibility of both recipient and theme was coded as ‘given’, and the animacy of
the recipient as ‘animate’. I then used the model to predict the argument realiza-
tion of the 210 sentences according to these variables. This procedure returns, for
each sentence, the probability that it is realized as one or the other variant. The
closer this probability score is to 1, the more likely is the to-dative variant to be
used; conversely, the closer it is to 0, the more likely is the ditransitive variant to
be used. I used a 0.5 threshold to determine the form predicted by the model: if
the probability score exceeds 0.5, the predicted form is to-dative, otherwise the
predicted form is ditransitive. I then compared the predictions of the model with
the actual variant used in each sentence.
As it turns out, there is a mere 52% match between the model’s predictions and
the actual realizations. However, the matches are unevenly distributed among the
two verb classes. If the two kinds of verbs are analyzed separately, it turns out that
72% of the productions with verbs of physical transfer are correctly predicted, versus
only 35% for verbs of communication. In sum, the factors influencing the choice of
variant in the dative alternation in actual discourse do seem to accurately predict
the variant chosen by subjects in a sizable portion of their productions (even more
so for verbs of physical transfer than for verbs of communication). Yet, it remains
to be seen whether the asymmetry reported above could be accounted for by these
factors only, or whether the same trend is also found in production choices not
predicted by the model. To do so, I split the dataset of each semantic class into two
types of productions: those in which the variant actually used by the subject matches
3. The training data was downloaded from the book’s companion website at http://www.black-
wellpublishing.com/quantmethods/. Following the R code provided by Johnson (2008, p.â•›251), I
used the glm function of the R environment.
192 Argument Structure in Usage-Based Construction Grammar
the prediction of the model, and those in which the used variant does not match
the prediction of the model. I then compared the distribution of these two types of
productions in terms of conservative vs. productive uses in each model construction
condition. These distributions are diagrammed respectively in Figure 7.4 for verbs
of physical transfer, and in Figure 7.5 for verbs of communication.
1.0 1.0
% of dative productions
% of dative productions
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
Model: Model: Model: Model:
ditransitive to-dative ditransitive to-dative
Productions matching Productions not matching
the predictions (N = 72) the predictions (N = 28)
Figure 7.4╇ Distribution of productions matching vs. not matching the predictions
of the logistic regression model for verbs of physical transfer.
1.0 1.0
% of dative productions
% of dative productions
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
Model: Model: Model: Model:
ditransitive to-dative ditransitive to-dative
Productions matching Productions not matching
the predictions (N = 39) the predictions (N = 71)
Figure 7.5╇ Distribution of productions matching vs. not matching the predictions
of the logistic regression model for verbs of communication.
Chapter 7.╇ The usage basis of alternation-based productivity 193
For verbs of physical transfer, the distribution of productions matching the pre-
dictions of the model (on the left) clearly presents the productivity asymmetry
reported earlier. This means that this asymmetry can be predicted by contextual
factors alone. However, the asymmetry is also found in cases that are not predicted
by the model (on the right), albeit to a lesser extent, since in the ditransitive presen-
tation condition, subjects were equally likely to produce either variant. However,
they also display a conservative behavior in the to-dative variant, which is what the
asymmetry primarily amounts to. In sum, if subjects were indeed sensitive to the
factors governing the alternation of dative verbs in their productions, their compli-
ance with such factors only reinforced an effect also found when their productions
appear to violate them, but is not solely responsible for this effect. In the latter case,
their behavior thus seems to rest on “genuine” productivity or conservativeness,
not motivated by contextual factors.
A quite different picture is found for verbs of communication, in that a differ-
ent construction is prevalent in each set of productions: respectively the ditransi-
tive variant for the productions that match the predictions of the model, and the
to-dative variant for those that do not. The fact that the model mostly predicts
ditransitives entails that the restrictions on the dative alternation should have dis-
couraged subjects from using the to-dative, which means that the other 65% of
productions in which they did use this variant were genuine cases of conservative-
ness or productivity. In other words, the possible tendency to follow the restrictions
on the alternation was in most cases outweighed by other factors, resulting in the
asymmetry previously reported. It also means that the apparently lesser strength
of the asymmetry for verbs of communication than for verbs of physical transfer
might be precisely due to the fact that the ditransitive variant is a more natural
choice when the factors governing the dative alternation are taken into account.
Be that as it may, the core of my initial observation remains: there is a productivity
asymmetry with verbs of communication.
In conclusion, the factors that govern the choice of variant for alternating da-
tive verbs in language use, as documented by Bresnan et al. (2007), fall short of
explaining the asymmetry found in the dative alternation. In the case of verbs of
physical transfer, these factors might make the asymmetry appear stronger than
it actually is, but do not seem to be solely responsible for it. In the case of verbs of
communication, their predictions run counter to the asymmetry and only account
for a minority of cases. All in all, the conservativeness of subjects in the to-dative
condition, and their more productive behavior in the ditransitive condition, ap-
pear to hold regardless of whether such factors might have played a role. The only
finding that should be taken with a note of caution is the varying strength of the
effect according to the verb class.
194 Argument Structure in Usage-Based Construction Grammar
7.2.4â•… Summary
In this part of the chapter, I presented evidence for the existence of productivity
asymmetries in alternations. I reported on an experiment in which subjects were
asked to produce a sentence with a novel verb that was previously presented to them
in one or the other variant of the dative or locative alternation. I found that there is
a productivity asymmetry in the dative alternation with adult speakers, similar to
the one Conwell & Demuth (2007) found with children. This asymmetry consists
in a more conservative behavior with verbs presented in the to-dative construction
than with verbs presented in the ditransitive construction. I also found that this
asymmetry holds both with verbs of physical transfer and with verbs of communica-
tion. As to the locative alternation, I did not find any such difference in productivity
between the two presentation conditions, irrespective of verb meaning.
In conclusion, this experiment provides evidence for cases of syntactic produc-
tivity where the construction with which a verb is first presented has an influence
on the probability for that verb to be used in another construction. This finding
seems at odds with a model of productivity solely driven by the meaning of con-
structions. In the next section, I show that it receives a better usage-based explana-
tion in a model taking alternation-based generalizations into account, giving some
credence to the hypothesis of alternation-based productivity.
In this section, two kinds of explanations are suggested for the productivity asym-
metry found in the dative alternation, and the lack of asymmetry in the locative
alternation. The first kind is based on contextual factors influencing the choice
of a variant of the dative alternation, and as such is only meant to account for
productivity in this alternation. Since the contextual factors governing the da-
tive alternation were not systematically controlled for, I reason that, if the novel
verbs were presented in a context that actually favors the other variant, subjects
could have inferred that the verb could not be used in this other variant, in line
with the idea of constructional preemption suggested in the literature. However,
I find that this explanation has only limited predictive power. The second kind of
explanation, more in line with the spirit of this book, is based on the usage of the
relevant constructions, especially in terms of type frequency. I report from a cor-
pus study that there are much fewer verbs in general English usage that occur in
both dative constructions than verbs than occur only in the to-dative; conversely,
almost all verb occurring in the ditransitive construction also occur in the to-
dative. This imbalance in type frequency correlates with the asymmetry found in
Chapter 7.╇ The usage basis of alternation-based productivity 195
As already mentioned, while the ditransitive construction and the to-dative construc-
tions are essentially synonymous, they still differ in terms of a number of contextual
restrictions. In any given instance, factors of discourse accessibility, pronominality,
definiteness, length and animacy of the recipient and theme arguments restrict (or
at least strongly bias) the use of alternating verbs to one particular variant. Yet, if one
of the variants is not permitted for a given verb, the allowed construction must be
used even in a discourse context that would in principle disfavor it. Goldberg (2011,
p.â•›144 and footnote 5) illustrates this point with the following pairs of examples:
(7) a. She explained the problem to me.
b. #She told the problem to me.
(8) a. She returned the shirt to me.
b. #She gave the shirt to me.
(9) a. They transferred the film rights to her.
b. #They sent the film rights to her.
The non-alternating verbs explain, return and transfer are perfectly fine in a context
that would favor the ditransitive construction (pronominal recipient, lexical theme
longer than recipient), whereas the same examples with the alternating verbs tell, give
and send sound pragmatically ill-formed. Hence, the contextual restrictions on the
to-dative construction do not equally apply to alternating vs. non-alternating verbs.
The idea that differences in the pragmatic behavior of verbs are connected to
their alternation behavior with respect to several related constructions is a key com-
ponent in the process of statistical preemption, hypothesized to explain how speak-
ers acquire arbitrary restrictions on the distribution of constructions by deriving
a form of indirect negative evidence from language use (cf. Goldberg, 2006). The
basic rationale of statistical preemption is that the inability of a word to occur in a
construction A can be inferred from the repeated exposure to the use of that word in
another construction B in contexts where using A would have been more felicitous.
For example, in the case of the dative alternation, speakers would be able to infer
that a verb like explain can only be used in the to-dative construction after repeatedly
witnessing this verb with the to-dative construction in contexts that would actually
favor the ditransitive (e.g., with a pronominal recipient and a lexical theme).
196 Argument Structure in Usage-Based Construction Grammar
In my experiment, the novel verbs were presented only once to subjects, thus it
is not clear whether statistical preemption really applies, since it is rather conceived
of as operating on the basis of repeated instances. Yet, if the single occurrence of the
novel verb involved a mismatch between the discourse context and the construc-
tion with which the verb was used, it is still possible that subjects were reluctant
to assume that the verb could be used with the other construction, along the lines
of statistical preemption. In other words, if the critical stimuli sentences had a
context favoring the ditransitive, this variant might have been preempted when the
to-dative was used instead, resulting in the more conservative behavior of subjects
in this presentation condition. To test this hypothesis, I used the logistic regression
model described in Section 7.2.3 to generate prediction scores according to the
properties of the theme and recipient arguments in the stimuli sentences contain-
ing novel verbs. The results of this analysis are presented in Table 7.7. As before,
the predicted form is to-dative if the prediction score is above the 0.5 threshold,
and ditransitive if it is below 0.5; the predicted form is reported in the last column.
The ID numbers found in the first column correspond to the numbering of verb
meanings listed in Section 7.1.3.
Table 7.7╇ Prediction scores and predicted form for the dative critical sentences.
ID Theme Recipient P(to-dative) Predicted form
Verbs of physical transfer
1 a very heavy box Ted 0.1 ditransitive
2 food supplies the citizens 0.55 to-dative
3 a package Susan 0.39 ditransitive
4 a letter Lisa 0.39 ditransitive
Verbs of communication
1 the news Stella 0.16 ditransitive
2 the shocking news the citizens 0.44 ditransitive
3 his plan Laura 0.16 ditransitive
4 the location of a source of food an ant 0.19 ditransitive
As it turns out, all critical sentences but one present a context that favors the di-
transitive variant, according to the predictions of the model. Hence, the to-dative
versions of these sentences do not match the restrictions of the to-dative con-
struction, which seems to confirm the hypothesis that subjects assumed the verb
did not alternate because of contextual factors. However, it should be noted that
for four sentences, the model does not make clear predictions, as the prediction
scores (marked in boldface font in Table 7.7) are closer to 0.5 than to 0 or 1. This
means that even if a strict application of the 0.5 threshold predicts a ditransitive or
Chapter 7.╇ The usage basis of alternation-based productivity 197
to-dative form, the other variant is not strikingly infelicitous. In other words, both
variants are nearly as acceptable, which is confirmed by introspective judgments:
note, for example, that Jack sent a package to Susan does not sound pragmatically
ill-formed compared to the predicted ditransitive equivalent Jack sent Susan a
package. In sum, these sentences do not seem to contain contextual cues strong
enough for subjects to reject the assumption that the verb may also be used in the
ditransitive construction, especially on the basis of a single instance. Additionally,
I should note that no noticeable difference in productivity is found between the
sentences that receive clear predictions and those that receive more uncertain ones.
In conclusion, even though it appears that contextual factors could have played a
role, they do not seem to provide a comprehensive explanation of the productivity
asymmetry in the dative alternation.
Concerning the locative alternation, it is not clear whether contextual factors
related to the alternating arguments play a role in the choice of variant, as it was
shown for the dative alternation. According to Laffut (1997), the choice between
variants in the locative alternation cannot be attributed solely to information struc-
ture considerations; instead, Laffut argues that affectedness is the critical factor,
which is congruent with the position I took in the previous chapter, namely that
the semantic difference between each locative variant is best defined in terms
of construal, with the locative variant construing the action of the agent as one
exerted on the theme, and the with-variant construing it as an action on the loca-
tion. Hence, the choice between variants stems from the properties of the event
itself (if one construal lends itself better to the description of that event than the
other), or from a deliberate semantic choice on behalf of the speaker (if both con-
struals are possible). In my experiment, I designed novel locative verb meanings
that were either container-oriented or content-oriented: the former type makes
reference to what happens to the location, while the latter type rather focuses on
what kind of motion is imparted to the theme; this contrast was reflected in the
definitions provided to subjects. There was however no difference in productivity
between these verb classes, which renders a semantic explanation rather dubious.
Yet, as I am about to show, the lack of asymmetry in the locative alternation can
be explained in usage-based terms, along the same lines as the asymmetry found
in the dative alternation.
Gropen et al. (1989) […] showed that a greater number of (lower frequency) verb
types actually occur in the prepositional form of the dative. The reasons for this
distributional fact are unclear, but the type/token imbalance suggests that children
may have some evidence that not all verbs can occur in the double object form or
that they do so only very rarely. (p.â•›176)
One way to formulate Conwell & Demuth’s (2007) suggestion in usage-based terms
is by referring to the type frequency of syntactic constructions, i.e., the number
of different items occurring in them. A large body of research has documented
the role of type frequency as a major determinant of productivity (Barðdal, 2008;
Bybee, 1995; Bybee & Thompson, 1997; Perek, 2014a; to appear; Suttle & Goldberg,
2011; Wonnacott et al., 2012; Zeschel, 2012), which is summarized by Goldberg
(2006) as follows:
Constructions that have appeared with many different types are more likely to
appear with new types than constructions that have only appeared with few types.
For example, argument structure constructions that have been witnessed with
many different verbs are more likely to be extended to appear with additional
verbs. To some extent, this observation has to be correct: a pattern is considered
extendable by learners only if they have witnessed the pattern being extended.
(p.â•›99)
In a nutshell, the higher the existing type frequency of a construction, the more
likely will this construction attract new members. There are two ways in which
type frequency could explain the results of my experiments, depending on whether
the variants of each alternation are considered as unrelated constructions or as
allostructions, as argued in Chapter 6. If the two variants of the alternation are
considered as independent constructions and not explicitly related in the gram-
mar, then we should only expect effects of the type frequency of constructions. In
other words, if one of the variants of the alternation has a higher type frequency
than the other variant, then the former should be more productive than the lat-
ter, and therefore, should be preferred by participants ceteris paribus. If, however,
variants of an alternation are considered as allostructions, then productivity could
depend not only on the type frequency of each allostruction but also on the type
frequency of the constructeme, i.e., on the number of alternating verbs. More spe-
cifically, if many verbs already alternate, participants should be more likely to as-
sume that novel verbs alternate too. Conversely, if there are many non-alternating
verbs, participants should be more likely to use verbs conservatively in the relevant
construction. This is, in fact, precisely what Wonnacott et al. (2008) find in their
artificial language learning study: the more alternating verbs were witnessed by
participants in the artificial language, the more likely they were to use verbs (both
known and novel) productively (cf. Section 6.4).
Chapter 7.╇ The usage basis of alternation-based productivity 199
In the case of the dative alternation, Conwell & Demuth (2007) seem to argue
against a type-frequency-based explanation, on the grounds that their children
subjects were not likely to have acquired the many Latinate verbs that constitute the
bulk of to-dative-only verbs. However, they observe that “some verbs, particularly
those of long-distance transfer (e.g., throw), are not commonly used in the double
object form” (p.â•›176), which would explain that children stick to the to-dative
construction when using novel verbs of this class. This explanation does rely to
some extent on type frequency, but instead of taking into account the whole set of
verbs that may be used in the relevant constructions, it considers only those verbs
that are actually used in them. Such a usage-based conception of type frequency
should be derived from corpus data; in other words, a construction that occurs
with many types in a given sample of language use is to be considered more pro-
ductive than another construction that occurs with fewer types in the same sample.
In the remainder of this section, I explore whether such an operationalization of
type frequency can explain the results of my experiment.
In order to do so, I looked at the distribution of dative and locative construc-
tions in the ICE-GB corpus. While this corpus contains only one million words,
it is fully annotated for syntactic structure, which makes it possible to easily ob-
tain the entire distribution of the relevant constructions in the whole corpus with
maximal precision. For the ditransitive uses, I queried the corpus for all clauses
marked “ditransitive” (either active or passive) and containing a nominal direct
object. I checked whether the verbs matched the central meaning of ‘transfer’ of the
ditransitive construction; in fact they all did to varying extents. For the to-dative
uses, I queried the corpus for all active and passive clauses containing a clause-level
prepositional phrase headed by to; for active clauses, I added the restriction that
the clause also had to contain a nominal direct object. I then manually filtered out
the sentences in which the to-phrase either is an adjunct, as in Example (10) below,
or does not correspond to a dative use, as in Example (11).
(10) We all do it to some degree, by accident sometimes. (W2D-017:102)
(11) We are also able to compare it to contemporary gatherer-hunters (W1A-011:122)
To qualify as a dative use, the clause has to describe an event of transfer, in which
the direct object refers to the transferred object, message, information, etc. (the
theme), and the to-phrase refers to the recipient of the theme. In most cases, this
translated into the requirement that the referent of the to-phrase be animate (since
it is a defining feature of recipients), but I also allowed inanimate referents in
case they were recipients in a metaphorical sense, as in Example (12) below. My
criterion of inclusion for metaphorical uses was whether the same verb could also
allow an animate referent for its to-phrase that can be categorized as a recipient.
200 Argument Structure in Usage-Based Construction Grammar
(12) In September 1921, the British newspapers gave prominence to one story
(W2B-019:3)
The usage data gathered from the ICE-GB is remarkably in line with Conwell &
Demuth’s (2007) comments concerning the usage of verbs of long-distance trans-
fer. Most verbs occurring exclusively in the to-dative do so with very low token
frequency, and nearly half of these verbs are actually hapax legomena, i.e., types
with a token frequency of one, which is also an index of high productivity (Baayen,
1992). I noticed, however, that many of these low-frequency verbs occurred in a
particular part of the corpus containing transcripts of sport commentaries. This is
not surprising, because commentators are likely to describe what happens during
ball games using adequate verbs of physical transfer; the particularly high frequency
of play, used in the sense of strike (a ball), is precisely contributed (for the most part)
by commentators’ speech. They also use a range of other verbs, some of which are
Chapter 7.╇ The usage basis of alternation-based productivity 201
part of a rather “technical” vocabulary (like lob and field). Others (like scoop and
spread) seem to be innovations of that register (if not coinages) and would prob-
ably not be considered as describing actions done with a ball by most speakers. In
sum, it appears that this part of the corpus unduly inflates the type frequency of
the construction. However, even if the sentences found in these texts are removed
from the distribution, there still remain twenty-eight verbs that occur exclusively in
the to-dative, which does not substantially change the overall picture. In summary,
there are many more verbs of physical transfer that occur exclusively in the to-dative
than verbs that occur in both the to-dative and the ditransitive in this sample of
English. On the other hand, it is also true that to-dative uses are significantly more
frequent than ditransitive uses for this class of verbs (267 vs. 74; χ2(1)â•›=â•›109.2346,
pâ•›<â•›0.001); it might therefore be the case that the strong association of this class with
the to-dative is an effect of token frequency, not type frequency.
A similar picture is obtained for the communication class. Only three verbs are
found to alternate in the corpus: tell, read and write; their frequency of occurrence
is reported in Table 7.10. Only one verb, ask, occurs exclusively in the ditransitive
(27 times); however, even though events of asking do involve communication, this
verb is rather peripheral to the category, since it does not describe the transfer of a
message but rather a request for information or action. By way of contrast, no less
than seventeen different verbs of communication are used only in the to-dative;
their distribution is reported in Table 7.11. As with verbs of physical transfer, the
distribution contains many hapax legomena. Again, there are more verbs occur-
ring in the to-dative construction only than in both dative constructions (17 vs. 3).
202 Argument Structure in Usage-Based Construction Grammar
Additionally, and contrary to what was found with verbs of physical transfer, verbs
of communication are used significantly more often in the ditransitive than in the
to-dative (137 vs. 103; χ2(1)â•›=â•›4.8167, pâ•›=â•›0.028). If token frequency really played a
role in the productivity of dative constructions, subjects would be expected to use
these verbs more frequently in the ditransitive than in the to-dative; but such was
not the case in my experiment. Hence, token frequency does not seem to account
for the asymmetry found in the dative alternation.
The type frequencies of the dative constructions are summarized in Table 7.12.4
In both semantic classes under consideration, most verbs occur exclusively in the
to-dative, much fewer verbs occur in both the ditransitive and the to-dative, and
even fewer verbs (only one in both class) occur in the ditransitive only. It should
be noted that this pattern of type frequencies is not unique to these two classes, but
is equally true of dative verbs as a whole. For reasons of space, I do not include the
full distribution of the ditransitive and to-dative uses, but I do provide their type
and token frequency in the last line of Table 7.12.
4. The number in brackets in the rightmost column is the remaining type frequency of verbs
of physical transfer when the sport commentaries texts are removed.
Chapter 7.╇ The usage basis of alternation-based productivity 203
The fact that there are more verbs occurring only in the to-dative construction than
in both constructions entails that, if all that is known about a given verb is that
it can be used in the to-dative, it is more likely that this verb belongs to the non-
alternating class than to the alternating one. The opposite holds for a novel verb
presented in a ditransitive use: since most dative verbs occurring in the ditransitive
also occur in the to-dative, there is little reason to assume (from a usage-based
point of view) that a verb used in the ditransitive cannot be used in the to-dative.
On the other hand, it is also the case that there are simply many more verbs attested
in the to-dative construction than in the ditransitive construction (at least 33 vs.
6 for physical transfer, 20 vs. 4 for communication, and 141 vs. 70 for all verbs).
Hence, even if we do not hypothesize a generalization over the dative variants of
alternating verbs (the dative constructeme), and thus a class of alternating verbs,
there is still a type frequency difference between the to-dative construction and the
ditransitive construction which predicts that the latter should be more productive
than the former, which can explain the overall preference of participants for the
to-dative variant without relying on the notion of alternation-based productivity.
In sum, both an account based on the type frequency of individual constructions
and one based on the size of alternating vs. non-alternating classes predict the
productivity asymmetry in the dative alternation found in the experiment.
However, the evidence from the locative alternation can help decide between
these two accounts. No asymmetry was found in the experiment in the case of the
locative alternation; rather, subjects were conservative with both constructions.
An explanation in terms of type frequency is also available for this behavior. I ex-
tracted all instances of the caused-motion construction and the with-applicative
construction from the ICE-GB corpus. For the former, I queried the corpus for
clauses containing a prepositional phrase headed by a locative preposition,5 and a
direct object if active. For the latter, I used the same query except that I restricted
the prepositional phrase to be headed by with. I filtered out the sentences that
5. I restricted the search to prepositions primarily used in a locative sense, leaving aside those
for which the prepositional complement is a source: above, across, against, along, around, behind,
below, between, beyond, down, down to, in, inside, into, near, on, onto, outside, over, past, round,
through, to, towards, under, up, up to, upon, within.
204 Argument Structure in Usage-Based Construction Grammar
the theme and recipient arguments (inter alia, cf. Section 6.3.2). The locative con-
structions, however, mostly differ in terms of construal, i.e., the perspective that
they adopt on the event of caused change of location, as argued in Section 6.3.3:
respectively an action on the theme vs. an action on the location. It is not entirely
clear to what extent construal (as opposed to propositional content) constrains the
productivity of a construction, but it is certainly more likely to “block” verbs from
occurring in it than pragmatic restrictions, for instance if the construal conveyed
by the construction clashes with the meaning of the verb. Hence, it could be the
case that participants were simply reluctant to apply an alternative construal to
a locative verb they had just witnessed, regardless of the original construal they
had witnessed it with. While I agree that explanations of this kind are generally
relevant to productivity, there are good reasons to think that it cannot alone
entirely account for the behavior of participants in this particular case. First, all
of the novel verb meanings make reference to both a particular way of moving/
applying the theme and a particular effect that may be brought about on the
location. Many accounts of the locative alternation revolve (to varying extents)
around the idea that it is precisely the presence of these two semantic components
that allow locative verbs to alternate (Pinker, 1989; Nemoto, 2005; Iwata, 2005,
2008). Hence, nothing clearly makes the novel verbs semantically incompatible
with either construction; in fact, both variants of the stimuli sentences do sound
perfectly fine (see Appendix). Second, it was found that productivity did not vary
according to whether the definition selected by subjects in the meaning decision
task (which occurred just before the sentence production task) presented the verb
as content-oriented or container-oriented (cf. Figure 7.3, p.â•›192). This finding
seems at odds with an explanation based on the compatibility of verb meanings
with different construals: since a content-oriented definition emphasizes the mo-
tion component of the verb’s meaning and a container-oriented definition the
effect component, it is expected that the caused-motion construction should be
preferred with the former and the with-applicative construction with the latter. As
already mentioned, this expectation is not borne out, which undermines the idea
that semantic factors alone could account for the generally conservative behavior
of participants in the locative alternation.
In sum, these results suggest that it is necessary to posit classes of alternating
verbs to account for the behavior of participants with respect to both the dative
and locative alternations, which points to the importance of alternation-based
generalizations in a construction grammar account of syntactic productivity. In
the next section, after recapitulating the findings of this chapter, I return to the
implications of this result.
206 Argument Structure in Usage-Based Construction Grammar
in which the minimal-exposure verb was presented, even though there was a type
frequency difference between the two constructions (in other words, they did not
generalize the verb to the construction with the highest type frequency). In the
generalist language, verbs were readily generalized to the unattested construction.
Wonnacott (2011) replicate these results with 6-year-olds, and Perek & Goldberg
(to appear) report a similar effect (albeit of a lower magnitude) with a pair of
constructions that differ in pragmatic function. Wonnacott et al. argue that the
relevant factor is “the type frequency of a shared distributional pattern, rather than
of a particular linguistic structure” (p.â•›205), which bears striking resemblance to
my own conclusions regarding the usage basis of productivity in the dative and
locative alternations, where only a comparison of the number of verbs occurring in
either one construction vs. both constructions appeared to provide an explanation
for the asymmetry of the alternation, or the absence thereof. In fact, this chapter
provides a response to Wonnacott et al.’s call to “explore […] corpora to ascertain
how the factors identified in these experiments apply to different natural language
constructions and how these statistical patterns line up with the varying degrees of
generalization shown by native speakers for different structures” (ibid.).
To conclude this chapter, I would like to discuss how my results relate to an-
other approach to argument structure productivity in the usage-based literature:
statistical preemption. As previously mentioned, the process of statistical preemp-
tion is aimed at explaining how speakers acquire arbitrary restrictions on the dis-
tribution of constructions, by deriving a form of indirect negative evidence from
language use, and rests on the assumption that the inability of a word to occur in
a construction A can be inferred from the repeated exposure to the use of that
word in another construction B in contexts where using A would have been more
felicitous. An obvious way in which the idea of statistical preemption relates to
my proposal to consider alternations as units of linguistic knowledge is that, as
mentioned in Chapter 6, it presupposes that speakers extract equivalence relations
between constructions and activate such representations in language comprehen-
sion; the allostructions model appears to be well tailored to this task. Apart from
this, it would seem as if statistical preemption has little to do with the present ap-
proach, because I am concerned with novel verbs witnessed by subjects only once,
whereas statistical preemption operates on the basis of repeated uses. However,
there are two ways in which the two accounts could inform each other.
First, I found that verbs are not as readily generalized to other constructions
as one might expect, and that this depends on the construction with which a verb
is previously witnessed, which I explained in terms of corresponding asymmetries
in type frequency. As it were, statistical preemption might thus play only a minor
role in those cases where strong predictions on productivity can already be made
on the sole basis of statistical relations between constructions, which means that
208 Argument Structure in Usage-Based Construction Grammar
the properties of the whole system of constructions should be taken into account
before gauging the role of statistical preemption.
On the other hand, there is also a way in which statistical preemption might
itself shape these properties. Originally, statistical preemption was thought to oper-
ate with individual items; however, in a recent experiment, Boyd & Goldberg (2011)
found that speakers are able to transfer knowledge derived from statistical preemp-
tion to similar items. An alternative explanation for the reluctance of subjects to
use to-dative verbs productively in the ditransitive construction might therefore be
that many verbs were statistically preempted from that construction in their past
linguistic experience. Under this account, the novel to-dative verbs would be as-
similated to an existing category of to-dative-only verbs (presumably on the basis
of syntactic and semantic similarity), with the knowledge that such verbs cannot be
used in the ditransitive construction. In sum, it remains to be seen to what extent
the present model of productivity relates to statistical preemption, and especially
whether statistical preemption is needed for a productivity asymmetry to appear,
or whether a type frequency imbalance alone is responsible for this effect.
Conclusion
chapter 8
The goal of this monograph was to investigate the usage basis of verb argument re-
alization. As noted in Chapter 1, verbs are not uttered in isolation: their use requires
reference to the participants involved in the scene they describe (i.e., their argu-
ments). Over the past decades, research on this topic has clearly shown that a theory
of argument realization cannot simply amount to a listing of the combinatorial
properties of each verb and that there are broader principles underlying argument
realization that must be an integral part of the linguistic knowledge of speakers.
The usage-based approach takes the view that grammar is a dynamic inven-
tory of symbolic conventions that is shaped by actual language use. Many studies
have substantiated the claim that the structure of grammar is ultimately tied to
usage, and no less may be said about that particular area of grammar concerned
with argument realization. Yet, accounts of the usage basis of argument realization
are rather piecemeal, and while many scholars assume some degree of relatedness
between verbal usage and argument realization, they often say little about what
this relation exactly consists in. This book sought to mend at least part of this gap.
In this section, I evaluate to what extent this goal was achieved.
I started with the lowest level that a theory of argument realization must be
concerned with: verbs themselves. The main research question in this part was to
determine the amount of argument realization information that is stored at the
level of verbs. In Chapter 2, I contrasted two broad families of approaches to argu-
ment realization: projectionist approaches and constructional approaches. These
two kinds of theories make radically different assumptions about the role of the
verb as a unit of argument realization information. In projectionist approaches,
argument realization entirely depends on verbs, and is usually formulated as rules
predicting the realization of the arguments of a verb from particular aspects of its
meaning. However, this verb-centric position is challenged by the fact that many
verbs can be used with several different argument structures, which can only be
captured by positing different meanings of verbs corresponding to their various
argument structures, resulting in rampant (and somewhat ill-grounded) verbal
polysemy. In addition, a projectionist approach is also undermined by the fact
that verbs can be used innovatively in atypical syntactic environments, for which
positing a new meaning of the verb is clearly counter-intuitive. The solution put
212 Argument Structure in Usage-Based Construction Grammar
(seller) vs. for (money) for buy. I then compared the results of this experiment to
frequency data collected from corpora of conversational British and American
English, in order to test whether the differences in cognitive accessibility found
in the experiment correlate with differences in frequency. I found that for sell and
pay, the most rapidly integrated valency pattern in the maze task (respectively
seller-goods-buyer, as exemplified by Bill sold his bike to the neighbor, and
buyer-money-goods, as exemplified by Laura paid forty euros for the meat) is in
each case more frequent in both corpora than the less rapidly integrated valency
pattern (respectively seller-goods-money, as exemplified by Bill sold his bike for
seventy euros, and buyer-money-seller, as exemplified by Laura paid forty euros
to the butcher). For buy, I did not find any substantial difference in cognitive acces-
sibility between the valency pattern buyer-goods-seller, as exemplified by Peter
bought a camera from the department store, and the pattern buyer-goods-money, as
exemplified by Peter bought a camera for seventy euros, whereas there was indeed
a difference in frequency. Yet, since this difference was weaker in this latter case
than for sell and pay, I did not interpret this result as fundamentally invalidating
my hypothesis. In sum, my case study seems to provide evidence for the idea that
the cognitive status of valency is related to usage, as it shows that more frequent
valency patterns of a verb are more cognitively accessible.
In Chapter 4, I turned to the next level of argument realization: argument
structure constructions, defined as pairings of an abstract meaning containing a
set of arguments with syntactic information as to how these arguments are lin-
guistically realized. Argument structure constructions are the construction gram-
mar implementation of verb-general principles of argument realization, since the
constructional meaning conveyed by a construction determines the range of verbs
that may be used in that construction via general principles of semantic compat-
ibility, as well as the semantic interpretation of the resulting verb-argument struc-
ture combinations. I presented a substantial body of evidence from the literature
pointing to the usage basis of constructional meaning. Many studies report a close
connection between the meaning of a construction and the meaning of verbs that
typically occur in it, which suggests that constructional meaning is extracted from
lexical material. In this usage-based account, a construction is viewed as a corre-
lation between syntactic form and lexical meaning, emerging as a form-meaning
pair: the meaning of the verb or verb class most prominently occurring in a given
syntactic pattern comes to be associated with that pattern.
While this hypothesis seems to hold for many constructions, I argued that it
is probably too simplistic in its present formulation. I reported on the case of the
conative construction, for which the relation between the constructional mean-
ing and the meaning of typical verbs is much more tenuous than for the exam-
ples previously discussed in the literature. At the most general level, the conative
214 Argument Structure in Usage-Based Construction Grammar
realize their theme and recipient arguments: either as two post-verbal NPs (e.g.,
John gave Mary a book), or as a post-verbal NP and a prepositional phrase headed
by to (e.g., John gave a book to Mary). In constructional approaches, alternations
are usually not included in the grammar as such, and are instead merely seen as a
manifestation of the fact that some verbs are semantically compatible with two dis-
tinct constructions, without there being any explicit relation stored between these
constructions. This position mainly stems from a dedication to describe grammar
in terms of generalizations of a common meaning and form, thus disregarding
generalizations of a similar meaning over distinct forms. I took issue with this ap-
proach, by pointing out that a grammar ignoring semantic and syntactic relations
between constructions might not be a complete account of the syntactic knowledge
of speakers. Following a proposal by Cappelle (2006), I suggested that the variants
of an alternation should be viewed as allostructions, i.e., elaborations of a formally
underspecified higher-level construction, called the constructeme, associated with
the meaning shared by the variants of the alternation. I applied the allostructions
model to two well-known alternations in English: the dative alternation and the
locative alternation. I then presented empirical evidence from a sorting task (Perek,
2012) and two priming experiments (Goldwater et al., 2011; Vasilyeva & Waterfall,
2011) suggesting that speakers indeed store generalizations of a common meaning
over formally different constructions. Furthermore, I suggested that allostructions
might play a role in argument structure productivity, in that the occurrence of
some verb in a particular allostruction triggers the expectation that this verb can
also be used in the other allostruction(s). Under this account, the other construc-
tions with which a verb has been previously observed, and not only properties of
individual argument structure constructions, should play a role in determining
the productivity of argument structure.
In Chapter 7, I presented evidence for this claim. I reported on an experi-
ment in which subjects were presented with novel verbs in one of the variants of
the dative alternation or the locative alternation, and were then asked to use the
verb in a sentence. I found that the construction with which the verbs were first
presented does have an influence on whether subjects used the verb in the other
variant of the alternation, in that they were mostly conservative when the verb
was first presented in the to-dative construction, but not when it was presented in
the ditransitive construction. Moreover, I found that the behavior of subjects can
largely be explained by usage properties of the relevant constructions, in terms of
differences in type frequency. Since many more verbs occur only in the to-dative
construction than in both variants of the alternation, a new verb used in the to-
dative can be considered more likely to be restricted to this construction than to
alternate. Conversely, since there are fewer ditransitive-only verbs than alternat-
ing verbs, a new verb used in the ditransitive can be considered likely to alternate,
216 Argument Structure in Usage-Based Construction Grammar
and is readily used in the to-dative. The absence of asymmetry in the locative
alternation was explained along the same lines. In sum, the behavior of speakers
is explained by the usage of constructions, though not taken independently but,
crucially, with reference to one another, which suggests that speakers’ knowledge
of statistical patterns in language use does include information about relations
between constructions that enter into an alternation.
Akhtar, N., & Tomasello, M. (1997). Young children’s productivity with word order and verb
morphology. Developmental Psychology, 33(6), 952–965. DOI: 10.1037/0012-1649.33.6.952
Albright, A., & Hayes, B. (2003). Rules vs. analogy in English past tenses: A computational/
experimental study. Cognition, 90, 119–161. DOI: 10.1016/S0010-0277(03)00146-X
Allerton, D. (1982). Valency and the English verb. London/New York: Academic Press.
Anderson, S.â•›R. (1971). On the role of deep structure in semantic interpretation. Foundations
of Language, 7(3), 387–396.
Arnon, I., & Snider, N. (2010). More than words: Frequency effects for multi-word phrases.
Journal of Memory and Language, 62, 67–82. DOI: 10.1016/j.jml.2009.09.005
Baayen, H. (2008). Analyzing linguistic data: A practical introduction to statistics using R. �Cambridge:
Cambridge University Press. DOI: 10.1017/CBO9780511801686
Baayen, H., & Lieber, R. (1991). Productivity and English derivation: A corpus-based study.
Linguistics, 29, 801–844. DOI: 10.1515/ling.1991.29.5.801
Baayen, H., & Milin, P. (2010). Analyzing reaction times. International Journal of Psychological
Research, 3(2), 12–28.
Baayen, H., Davidson, D.â•›J., & Bates, D.â•›M. (2008). Mixed-effects modeling with crossed random
effects for subjects and items. Journal of Memory and Language, 59, 390–412.
DOI: 10.1016/j.jml.2007.12.005
Baayen, R.â•›H. (1992). Quantitative aspects of morphological productivity. In G.â•›E. Booij & J. van
Marle (Eds.), Yearbook of Morphology 1991 (pp. 109–149). Dordrecht: Kluwer Academic
Publishers. DOI: 10.1007/978-94-011-2516-1_8
Baker, C.â•›F., & Ruppenhofer, J. (2002). FrameNet’s Frames vs. Levin’s Verb Classes. In J. Larson
& M. Paster (Eds.), Proceedings of the 28th Annual Meeting of the Berkeley Linguistics Society
(pp. 27–38). Berkeley, CA: Berkeley Linguistics Society.
Baker, C.â•›L. (1979). Syntactic theory and the projection problem. Linguistic Inquiry, 10(4), 533–581.
Bannard, C., & Matthews, D. (2008). Stored word sequences in language learning: The effect of
familiarity on children’s repetition of four-word combinations. Psychological Science, 19(3),
241–248. DOI: 10.1111/j.1467-9280.2008.02075.x
Barðdal, J. (2008). Productivity: Evidence from case and argument structure in Icelandic. Â�Amsterdam:
John Benjamins. DOI: 10.1075/cal.8
Barlow, M., & Kemmer, S. (Eds.). (2000). Usage-based models of language. Stanford: CSLI Publications.
Barsalou, L.â•›W. (1983). Ad hoc categories. Memory and Cognition, 11(3), 211–227.
DOI: 10.3758/BF03196968
Barsalou, L.â•›W. (1992). Frames, concepts, and conceptual fields. In E. Kittay & A. Lehrer (Eds.),
Frames, fields, and contrasts: New essays in semantic and lexical organization (pp. 21–74).
Hillsdale, NJ: Lawrence Erlbaum Associates.
Bates, E., Bretherton, I., & Snyder, L. (1988). From first words to grammar: Individual differences
and dissociable mechanisms. New York: Cambridge University Press.
220 Argument Structure in Usage-Based Construction Grammar
Bencini, G.â•›M.â•›L., & Goldberg, A.â•›E. (2000). The contribution of argument structure construc-
tions to sentence meaning. Journal of Memory and Language, 43(4), 640–651.
DOI: 10.1006/jmla.2000.2757
Biber, D. (1993). Representativeness in corpus design. Literary and Linguistic Computing, 8(4),
243–257. DOI: 10.1093/llc/8.4.243
Blumenthal-Dramé, A. (2012). Entrenchment in usage-based theories: What corpus data do and
do not reveal about the mind. Berlin: Mouton de Gruyter. DOI: 10.1515/9783110294002
Boas, H.â•›C. (2003). A constructional approach to resultatives. Stanford: CSLI Publications.
Boas, H.â•›C. (2008). Determining the structure of lexical entries and grammatical constructions
in Construction Grammar. Annual Review of Cognitive Linguistics, 6, 113–144.
DOI: 10.1075/arcl.6.06boa
Boas, H.â•›C. (2010). Linguistically relevant meaning elements of English communication verbs.
Belgian Journal of Linguistics, 24, 54–82. DOI: 10.1075/bjl.24.03boa
Boas, H.â•›C. (2011a). A frame-semantic approach to syntactic alternations with build-verbs. In
P. Guerrero Medina (Ed.), Morphosyntactic alternations in English (pp. 207–234). London:
Equinox.
Boas, H.â•›C. (2011b). Coercion and leaking argument structures in Construction Grammar. Lin-
guistics, 49(6), 1271–1303. DOI: 10.1515/ling.2011.036
Boas, H.â•›C. (2014). Lexical and phrasal approaches to argument structure: Two sides of the same
coin. Theoretical Linguistics, 40(1–2), 89–112. DOI: 10.1515/tl-2014-0003
Bod, R. (1998). Beyond grammar: An experience-based theory of language. Stanford: CSLI Publications.
Borer, H. (2003). Exo-skeletal vs. endo-skeletal explanations: Syntactic projections and the lexicon.
In J. Moore & M. Polinsky (Eds.), The nature of explanation in linguistic theory (pp. 31–67).
Stanford: CSLI Publications.
Bowerman, M. (1982a). Evaluating competing linguistic models with language acquisition data:
Implications of developmental errors with causative verbs. Quaderni di Semantica, III, 5–66.
Bowerman, M. (1982b). Reorganizational processes in lexical and syntactic development. In
E. Wanner & L.â•›R. Gleitman (Eds.), Language acquisition: The state of the art (pp. 319–346).
New York: Academic Press.
Bowerman, M. (1988). The ‘no negative evidence’ problem: How do children avoid constructing
an overly general grammar? In J. Hawkins (Ed.), Explaining language universals (pp. 73–101).
Oxford: Basil Blackwell.
Boyd, J.â•›K., & Goldberg, A.â•›E. (2011). Learning what not to say: The role of statistical preemption
and categorization in a-adjective production. Language, 87(1), 55–83.
DOI: 10.1353/lan.2011.0012
Boyd, J.â•›K., Gottschalk, E.â•›A., & Goldberg, A.â•›E. (2009). Linking rule acquisition in novel phrasal con-
structions. Language Learning, 59(Suppl. 1), 64–89. DOI: 10.1111/j.1467-9922.2009.00536.x
Braine, M.â•›D.â•›S. (1963). The ontogeny of English phrase structure: The first phase. Language,
39(1), 1–14. DOI: 10.2307/410757
Bresnan, J. (1982). The mental representation of grammatical relations. Cambridge, MA: MIT
Press.
Bresnan, J. (2001). Lexical-Functional Syntax. Oxford: Blackwell Publishers.
Bresnan, J. (2007). Is knowledge of syntax probabilistic? Experiments with the English dative
alternation. In S. Featherston & W. Sternefeld (Eds.), Roots: Linguistics in search of its evi-
dential base (pp. 75–96). Berlin/New York: Mouton de Gruyter.
221
References
Bresnan, J., & Ford, M. (2010). Predicting syntax: Processing dative constructions in American
and Australian varieties of English. Language, 86(1), 168–213. DOI: 10.1353/lan.0.0189
Bresnan, J., & Nikitina, T. (2009). The gradience of the dative alternation. In L. Uyechi & L.-H. Wee
(Eds.), Reality exploration and discovery: Pattern interaction in language and life (pp. 161–184).
Stanford: CSLI Publications.
Bresnan, J., Cueni, A., Nikitina, T., & Baayen, H. (2007). Predicting the dative alternation. In
G. Boume, I. Kraemer, & J. Zwarts (Eds.), Cognitive foundations of interpretation (pp. 69–94).
Amsterdam: Royal Netherlands Academy of Science.
Briscoe, T., & Copestake, A. (1999). Lexical rules in constraint-based grammars. Computational
Linguistics, 25(4), 487–526.
Broccias, C. (2001). Allative and ablative at-constructions. In M. Andronis, C. Ball, H. Elston,
& S. Neuvel (Eds.), CLS 37: The main session. Papers from the 37th meeting of the Chicago
Linguistic Society (Vol. 1, pp. 67–82). Chicago: Chicago Linguistic Society.
Brooks, P.â•›J., & Tomasello, M. (1999). Young children learn to produce passives with nonce verbs.
Developmental Psychology, 35(1), 29–44. DOI: 10.1037/0012-1649.35.1.29
Bybee, J., & McClelland, J.â•›L. (2005). Alternatives to the combinatorial paradigm of linguistic the-
ory based on domain general principles of human cognition. In N.â•›A. Ritter (Ed.), The role
of linguistics in cognitive science. Special issue of The Linguistic Review, 22(2–4), 381–410.
Bybee, J. (1985). Morphology: A study of the relation between meaning and form. Amsterdam:
John Benjamins. DOI: 10.1075/tsl.9
Bybee, J. (1995). Regular morphology and the lexicon. Language and Cognitive Processes, 10(5),
425–455. DOI: 10.1080/01690969508407111
Bybee, J. (2000). The phonology of the Lexicon: Evidence from lexical diffusion. In M. Barlow &
S. Kemmer (Eds.), Usage-based models of language (pp. 65–85). Stanford: CSLI Publications.
Bybee, J. (2006). From usage to grammar: The mind’s response to repetition. Language, 82(4),
711–733. DOI: 10.1353/lan.2006.0186
Bybee, J. (2010). Language, usage and cognition. Cambridge: Cambridge University Press.
DOI: 10.1017/CBO9780511750526
Bybee, J. (2013). Usage-based theory and exemplar representations of constructions. In T. �Hoffmann
& G. Trousdale (Eds.), The Oxford Handbook of construction grammar (pp. 49–69). Oxford:
�
Oxford University Press.
Bybee, J., & Eddington, D. (2006). A usage-based approach to Spanish verbs of ‘becoming’. Lan-
guage, 82(2), 323–355. DOI: 10.1353/lan.2006.0081
Bybee, J., & Hopper, P. (Eds.). (2001). Frequency and the emergence of linguistic structure. �Amsterdam:
John Benjamins. DOI: 10.1075/tsl.45
Bybee, J., & Moder, C.â•›L. (1983). Morphological classes as natural categories. Language, 59,
251–270. DOI: 10.2307/413574
Bybee, J., & Scheibman, J. (1999). The effect of usage on degrees of constituency: The reduction
of don’t in English. Linguistics, 37, 575–596. DOI: 10.1515/ling.37.4.575
Bybee, J., & Slobin, D.â•›I. (1982). Rules and schemas in the development and use of the English
past tense. Language, 58, 265–289. DOI: 10.1353/lan.1982.0021
Bybee, J., & Thompson, S. (1997). Three frequency effects in syntax. Berkeley Linguistics Society,
23, 65–85.
Cappelle, B. (2006). Particle placement and the case for “allostructions.” Constructions, Special
Volume 1, 1–28.
222 Argument Structure in Usage-Based Construction Grammar
Casenhiser, D., & Goldberg, A.â•›E. (2005). Fast mapping between a phrasal form and meaning.
Developmental Science, 8(6), 500–508. DOI: 10.1111/j.1467-7687.2005.00441.x
Chang, F., Bock, K., & Goldberg, A.â•›E. (2003). Can thematic roles leave traces of their places?
Cognition, 90, 29–49. DOI: 10.1016/S0010-0277(03)00123-9
Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press.
Chomsky, N. (1981). Lectures on government and binding: The pisa lectures. Berlin/New York:
Mouton de Gruyter.
Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT Press.
Clahsen, H., & Rothweiler, M. (1992). Inflectional rules in children’s grammars: Evidence from
the development of participles in German. In Yearbook of Morphology 1992 (pp. 1–34).
Clark, E., & Clark, H. (1979). When nouns surface as verbs. Language, 55(4), 767–811.
DOI: 10.2307/412745
Clark, E.â•›V. (1978). Discovering what words can do. Chicago Linguistic Society, 14, 34–57.
Clifton, C., Frazier, L., & Connine, C. (1984). Lexical expectations in sentence comprehension.
Journal of Verbal Learning and Verbal Behavior, 23, 696–708.
DOI: 10.1016/S0022-5371(84)90426-2
Colleman, T. (2009). The semantic range of the Dutch double object construction: A collos-
tructional perspective. Constructions and Frames, 1, 190–221. DOI: 10.1075/cf.1.2.02col
Colleman, T. (2010). Beyond the dative alternation: The semantics of the Dutch aan-dative. In
D. Glynn & K. Fischer (Eds.), Quantitative methods in cognitive semantics: Corpus-driven
approaches (pp. 271–303). Berlin/New York: Mouton de Gruyter.
Colleman, T., & De Clerck, B. (2009). ‘Caused motion?’ The semantics of the English to-dative
and the Dutch aan-dative. Cognitive Linguistics, 20, 5–42. DOI: 10.1515/COGL.2009.002
Collins, P. (1995). The indirect object construction in English: An informational approach. Lin-
guistics, 33, 35–49. DOI: 10.1515/ling.1995.33.1.35
Connine, C., Ferreira, F., Jones, C., Clifton, C., & Frazier, L. (1984). Verb frame preferences:
Descriptive norms. Journal of Psycholinguistic Research, 13(4), 307–319.
DOI: 10.1007/BF01076840
Conwell, E., & Demuth, K. (2007). Early syntactic productivity: Evidence from dative shift.
Cognition, 103, 163–179. DOI: 10.1016/j.cognition.2006.03.003
Croft, W. (1998). Linguistic evidence and mental representations. Cognitive Linguistics, 9(2),
151–173. DOI: 10.1515/cogl.1998.9.2.151
Croft, W. (2001). Radical construction grammar: Syntactic theory in typological perspective. O � xford:
Oxford University Press. DOI: 10.1093/acprof:oso/9780198299554.001.0001
Croft, W. (2003). Lexical rules vs. constructions: A false dichotomy. In H. Cuyckens, T. Berg,
R. Dirven, & K.-U. Panther (Eds.), Motivation in language: Studies in honour of Günter Â�Radden
(pp. 49–68). Amsterdam: John Benjamins. DOI: 10.1075/cilt.243.07cro
Croft, W. (2012). Verbs: Aspect and causal structure. Oxford: Oxford University Press.
DOI: 10.1093/acprof:oso/9780199248582.001.0001
Croft, W., Taoka, C., & Wood, E.â•›J. (2001). Argument linking and the commercial transaction
frame in English, Russian and Japanese. Language Sciences, 23, 579–602.
DOI: 10.1016/S0388-0001(00)00037-1
Dang, T.â•›H., Kipper, K., Palmer, M., & Rosenzweig, J. (1998). Investigating regular sense extensions
based on intersective Levin classes. Proceedings of COLING-ACL, Montréal (pp. 293–299).
Davies, M. (2008). The corpus of contemporary American English: 450 million words, 1990–Â�present.
Available online at http://corpus.byu.edu/coca/
223
References
Desmet, T., & Gibson, E. (2003). Disambiguation preferences and corpus frequencies in noun
phrase conjunction. Journal of Memory and Language, 49, 353–374.
DOI: 10.1016/S0749-596X(03)00025-1
Desmet, T., De Baecke, C., Drieghe, D., Brysbaert, M., & Vonk, W. (2006). Relative clause attach-
ment in Dutch: On-line comprehension corresponds to corpus frequencies when lexical
variables are taken into account. Language and Cognitive Processes, 21(4), 453–485.
DOI: 10.1080/01690960400023485
Diessel, H. (2007). Frequency effects in language acquisition, language use, and diachronic
change. New Ideas in Psychology, 25, 108–127. DOI: 10.1016/j.newideapsych.2007.02.002
Dik, S.â•›C. (1989). The theory of functional grammar, Parts 1 & 2. Berlin: Mouton de Gruyter.
Dixon, R.â•›M.â•›W. (1991). A new approach to English grammar, on semantic principles. Oxford:
Oxford University Press.
Dowty, D.â•›R. (1978). Lexically governed transformations as lexical rules in a Montague Gram-
mar. Linguistic Inquiry, 9, 393–426.
Dowty, D.â•›R. (1991). Thematic proto-roles and argument selection. Language, 67(3), 547–619.
DOI: 10.1353/lan.1991.0021
Eddington, D., & Ruiz de Mendoza Ibáñez, F. (2010). Argument constructions and language
processing: Evidence from a priming experiment and pedagogical implications. In S. De
Knop, F. Boers, & A. De Rycker (Eds.), Fostering language teaching efficiency through cogni-
tive linguistics (pp. 213–238). Berlin: Mouton de Gruyter.
Ellis, N.â•›C., & O’Donnell, M.â•›B. (2011). Robust language acquisition – an emergent consequence
of language as a complex adaptive system. In L. Carlson, C. Hölscher, & T. Shipley (Eds.),
Proceedings of the 33rd Annual conference of the cognitive science society (pp. 3512–3517).
Austin: Cognitive Science Society.
Ellis, N.â•›C., & O’Donnell, M.â•›B. (2012). Statistical construction learning: Does a Zipfian problem
space ensure robust language learning? In J. Rebuschat & J. Williams (Eds.), Statistical learn-
ing and language acquisition (pp. 265–304). Berlin: Mouton de Gruyter.
Ellis, N.â•›C. (1996). Sequencing in SLA: Phonological memory, chunking and points of order.
Studies in Second Language Acquisition, 18, 91–126. DOI: 10.1017/S0272263100014698
Ellis, N.â•›C. (2002). Frequency effects in language processing: A review with implications for
theories of implicit and explicit language acquisition. Studies in Second Language Acquisi-
tion, 24(2), 143–188.
Ellis, N.â•›C., & Ferreira-Junior, F. (2009). Construction learning as a function of frequency, fre-
quency distribution, and function. The Modern Language Journal, 93(iii), 370–385.
DOI: 10.1111/j.1540-4781.2009.00896.x
Ellis, N.â•›C., O’Donnell, M.â•›B., & Römer, U. (2014). The processing of verb-argument construc-
tions is sensitive to form, function, frequency, contingency and prototypicality. Cognitive
Linguistics, 25(1), 55–98. DOI: 10.1515/cog-2013-0031
Emonds, J. (1972). Evidence that indirect object movement is a structure-preserving rule. Foun-
dations of Language, 8(4), 546–561.
Erman, B., & Warren, B. (2000). The idiom principle and the open choice principle. Text, 20,
29–62.
Erteschik-Shir, N. (1979). Discourse constraints on dative movement. In T. Givón (Ed.), Syntax
and semantics: Vol. 12. Discourse and syntax (pp. 441–467). New York: Academic Press.
Faber, P.â•›B., & Mairal Usón, R. (1999). Constructing a lexicon of English verbs. Berlin: Mouton de
Gruyter. DOI: 10.1515/9783110800623
224 Argument Structure in Usage-Based Construction Grammar
Faulhaber, S. (2011a). Verb valency patterns: A challenge for semantics-based accounts. Berlin:
Mouton de Gruyter. DOI: 10.1515/9783110240788
Faulhaber, S. (2011b). Idiosyncrasy in verb valency patterns. Zeitschrift für Anglistik und Ame-
rikanistik, 59(4), 331–346. DOI: 10.1515/zaa-2011-0405
Fellbaum, C. (Ed.). (1998). WordNet: An electronic lexical database. Cambridge, MA: MIT Press.
Ferreira, F., & Henderson, J. (1991). Recovery from misanalyses of garden-path sentences. Jour-
nal of Memory and Language, 30(6), 725–745. DOI: 10.1016/0749-596X(91)90034-H
Ferretti, T.â•›R., McRae, K., & Hatherell, A. (2001). Integrating verbs, situation schemas, and the-
matic role concepts. Journal of Memory and Language, 44, 516–547.
DOI: 10.1006/jmla.2000.2728
Fillmore, C.â•›J., & Atkins, B.â•›T. (1992). Towards a frame-based Lexicon: The semantics of RISK
and its neighbors. In A. Lehrer & E. Kittay (Eds.), Frames, fields and contrasts: New essays
in semantic and lexical organization (pp. 75–102). Hillsdale: Erlbaum.
Fillmore, C.â•›J., & Atkins, B.â•›T. (1994). Starting where the dictionaries stop: The challenge for
computational lexicography. In B.â•›T. Atkins & A. Zampolli (Eds.), Computational approaches
to the Lexicon (pp. 349–393). Oxford: Oxford University Press.
Fillmore, C.â•›J. (1968). The case for case. In E. Bach & R. Harms (Eds.), Universals in linguistic
theory (pp. 1–88). New York: Holt, Rinehart, and Winston.
Fillmore, C.â•›J. (1970). The grammar of hitting and breaking. In R.â•›A. Jacobs & P.â•›S. Rosenbaum
(Eds.), Readings in English transformational grammar (pp. 120–133). Waltham, MA: Ginn.
Fillmore, C.â•›J. (1977). Topics in lexical semantics. In R.â•›W. Cole (Ed.) Current issues in linguistic
theory (pp. 76–138). Bloomington: Indiana University Press.
Fillmore, C.â•›J. (1985). Frames and the semantics of understanding. Quaderni di Semantica, VI(2),
222–254.
Fillmore, C.â•›J. (1999). Inversion and constructional inheritance. In G. Webelhuth, J.-P. Koenig, &
A. Kathol (Eds.), Lexical and constructional aspects of linguistic explanation (pp. 113–128).
Stanford: CSLI Publications.
Fillmore, C.â•›J., & Kay, P. (ms.). Construction grammar (course reader). Berkeley: University of
California.
Fillmore, C.â•›J., Lee-Goldman, R.â•›R., & Rhodes, R. (2012). The framenet constructicon. In H.â•›C. Boas
& I.â•›A. Sag (Eds.), Sign-based construction grammar. Stanford: CSLI Publications.
Forster, K.â•›I. (2010). Using a maze task to track lexical and sentence processing. The Mental
Lexicon, 5(3), 347–357. DOI: 10.1075/ml.5.3.05for
Forster, K.â•›I., Guerrera, C., & Elliot, L. (2009). The maze task: Measuring forced incremental
sentence processing time. Behavior Research Methods, 41(1), 163–171.
DOI: 10.3758/BRM.41.1.163
Fried, M., & Östman, J.-O. (2004). Construction grammar: A thumbnail sketch. In M. Fried
& J.-O. Östman (Eds.), Construction grammar in a cross-language perspective (pp. 11–86).
Amsterdam: John Benjamins. DOI: 10.1075/cal.2.02fri
Gahl, S. (2008). Time and thyme are not homophones: The effect of lemma frequency on word
durations in spontaneous speech. Language, 84(3), 474–496. DOI: 10.1353/lan.0.0035
Gahl, S., & Garnsey, S.â•›M. (2004). Knowledge of grammar, knowledge of usage: Syntactic prob-
abilities affect pronunciation variation. Language, 80(4), 748–775.
DOI: 10.1353/lan.2004.0185
Gahl, S., Jurafsky, D., & Roland, D. (2004). Verb subcategorization frequencies: American English
corpus data, methodological studies, and cross-corpus comparisons. Behavior Research
Methods, 36(3), 432–443. DOI: 10.3758/BF03195591
225
References
GarcĭaVelasco, D. (2009). Innovative coinage: its place in the grammar. In C. Butler & J. Martĭn Arista
(Eds.), Deconstructing constructions (pp. 3–23). Amsterdam/Philadelphia: John Â�Benjamins.
DOI: 10.1075/slcs.107.03inn
García Velasco, D. (2011). The causative/inchoative alternation in functional discourse gram-
mar. In P. Guerrero Medina (Ed.), Morphosyntactic alternations in English. Functional and
cognitive perspectives (pp. 115–136). London/Oakville: Equinox.
Garnsey, S.â•›M., Pearlmutter, N.â•›J., Myers, E., & Lotocky, M.â•›A. (1997). The contribution of verb
bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of
Memory and Language, 37, 58–93. DOI: 10.1006/jmla.1997.2512
Gawron, J.-M. (1983). Lexical representations and the semantics of complementation. PhD thesis,
University of California, Berkeley.
Givón, T. (1984). Syntax: A functional-typological introduction. Volume I. Amsterdam: John
Benjamins. DOI: 10.1017/s0022226700010434
Givón, T. (1990). Syntax: A functional-typological introduction. Volume II. Amsterdam: John
Benjamins. DOI: 10.1017/s0022226700015322
Glynn, D. (2010). Testing the hypothesis: Objectivity and verification in usage-based cognitive
semantics. In K. Fischer & D. Glynn (Eds.), Quantitative methods in cognitive semantics:
Corpus-driven approaches (pp. 239–270). Berlin: Mouton de Gruyter.
Goldberg, A.â•›E. (1995). Constructions: A construction grammar approach to argument structure.
Chicago: University of Chicago Press.
Goldberg, A.â•›E. (2002). Surface generalizations: An alternative to alternations. Cognitive Linguis-
tics, 13(4), 327–356. DOI: 10.1515/cogl.2002.022
Goldberg, A.â•›E. (2006). Constructions at work: The nature of generalization in language. Oxford:
Oxford University Press.
Goldberg, A.â•›E. (2011). Corpus evidence of the viability of statistical preemption. Cognitive
Linguistics, 22(1), 131–153. DOI: 10.1515/cogl.2011.006
Goldberg, A.â•›E. (2013). Argument structure constructions versus lexical rules or derivational
verb templates. Mind & Language, 28(4), 435–465. DOI: 10.1111/mila.12026
Goldberg, A.â•›E. (2014). Fitting a slim dime between the verb template and argument structure
construction approaches. Theoretical Linguistics, 40(1–2), 113–135.
DOI: 10.1515/tl-2014-0004
Goldberg, A.â•›E., & Casenhiser, D.â•›M. (2006). Learning argument structure constructions. In
E.â•›V. Clark & B. Kelly (Eds.), Constructions in acquisition (pp. 185–204). Stanford: CSLI
Publications.
Goldberg, A.â•›E., Casenhiser, D.â•›M., & Sethuraman, N. (2004). Learning argument structure gen-
eralizations. Cognitive Linguistics, 15(3), 289–316. DOI: 10.1515/cogl.2004.011
Goldwater, M.â•›B., Tomlinson, M.â•›T., Echols, C.â•›H., & Love, B.â•›C. (2011). Structural priming as
structure-mapping: Children use analogies from previous utterances to guide sentence
production. Cognitive Science, 35, 156–170. DOI: 10.1111/j.1551-6709.2010.01150.x
Green, G. (1974). Semantics and syntactic regularity. Bloomington: Indiana University Press.
Gregory, M.â•›L., Raymond, W.â•›D., Bell, A., Fosler-Lussier, E., & Jurafsky, D. (1999). The effects of
collocational strength and contextual predictability in lexical production. Chicago Linguistic
Society, 35, 151–166.
Gries, S.â•›T. (2003). Multifactorial analysis in corpus linguistics: A study of particle placement.
London & New York: Continuum Press.
Gries, S.â•›T. (2005). Syntactic priming: A corpus-based approach. Journal of Psycholinguistic Re-
search, 34(4), 365–399. DOI: 10.1007/s10936-005-6139-3
226 Argument Structure in Usage-Based Construction Grammar
Gries, S.╛T. (2007). New perspectives on old alternations. In J.╛E. Cihlar, A.╛L. Franklin, & D.╛W. �Kaiser
(Eds.), Papers from the 39th regional meeting of the Chicago Linguistic Society: Vol. II, the panels
(pp. 274–292). Chicago: Chicago Linguistic Society.
Gries, S.â•›T., & Stefanowitsch, A. (2004). Extending collostructional analysis: A corpus-based
perspective on “alternations.” International Journal of Corpus Linguistics, 9, 97–129.
DOI: 10.1075/ijcl.9.1.06gri
Gries, S.â•›T., Hampe, B., & Schönefeld, D. (2005). Converging evidence: Bringing together experi-
mental and corpus data on the association of verbs and constructions. Cognitive Linguistics,
16(4), 635–676. DOI: 10.1515/cogl.2005.16.4.635
Gropen, J., Pinker, S., Hollander, M., Goldberg, R., & Wilson, R. (1989). The learnability and
acquisition of the dative alternation in English. Language, 65(2), 203–257.
DOI: 10.2307/415332
Gruber, J.â•›S. (1965). Studies in lexical relations. Doctoral dissertation, MIT, Cambridge, MA.
Gruber, J.â•›S. (1967). Look and See. Language, 43(4), 937–947. DOI: 10.2307/411974
Guerrero Medina, P. (2011). An antipassive interpretation of the English “conative alternation”:
Semantic and discourse-pragmatic dimensions. In P. Guerrero Medina (Ed.), Morphosyn-
tactic Alternations in English: Functional and Cognitive Perspectives (pp. 182–203). London:
Equinox.
Habert, B. (2001). Des corpus représentatifs: De quoi, pour quoi, comment? In Linguistique sur
corpus: Études et réflexions (pp. 11–58). Presses Universitaires de Perpignan.
Hall, B.â•›C. (1965). Subject and object in modern English. MIT.
Halliday, M.â•›A.â•›K. (1994). Introduction to functional grammar, 2nd ed. London: Edward Arnold.
Hare, M.â•›L., Ford, M., & Marslen-Wilson, W.â•›D. (2001). Ambiguity and frequency effects in
regular verb inflection. In J. Bybee & P. Hopper (Eds.), Frequency and the emergence of
linguistic structure (pp. 181–200). Amsterdam: John Benjamins. DOI: 10.1075/tsl.45.10har
Hare, M., McRae, K., & Elman, J.â•›L. (2003). Sense and structure: Meaning as a determinant of
verb subcategorization preferences. Journal of Memory and Language, 48, 281–303.
DOI: 10.1016/S0749-596X(02)00516-8
Harris, C.â•›L. (1998). Psycholinguistic studies of entrenchment. In J.-P. Koenig (Ed.), Discourse
and cognition: Bridging the gap (pp. 55–70). Stanford: CSLI Publications.
Hay, J. (2001). Lexical frequency in morphology: Is everything relative? Linguistics, 39, 1041–1070.
DOI: 10.1515/ling.2001.041
Helbig, G. (1992). Probleme der Valenz- und Kasustheorie. Tübingen: Niemeyer.
DOI: 10.1515/9783110938326
Herbst, T., & Uhrig, P. (2010). Erlangen valency patternbank. Available at: http://www.pattern-
bank.uni-erlangen.de/
Herbst, T. (1983). Untersuchungen zur Valenz englischer Adjektive und ihrer Nominalisierungen.
Tűbingen: Narr.
Herbst, T. (2009). Valency – item-specificity and idiom principle. In U. Römer & R. Schulze
(Eds.), Exploring the lexis–grammar interface (pp. 49–68). Amsterdam: John Benjamins.
DOI: 10.1075/scl.35.05her
Herbst, T. (2010). Valency constructions and clause constructions or how, if at all, valency gram-
marians might sneeze the foam off the cappuccino. In H.-J. Schmid & S. Handl, (Eds.),
Cognitive foundations of linguistic usage patterns: Empirical studies (pp. 225–255). Berlin/
New York: Mouton de Gruyter.
Herbst, T. (2011). The status of generalizations: Valency and argument structure constructions.
Zeitschrift für Anglistik und Amerikanistik, 59(4), 331–346. DOI: 10.1515/zaa-2011-0406
227
References
Herbst, T., & Schüller, S. (2008). An introduction to syntactic analysis: A valency approach. TÂ� übingen:
Narr.
Herbst, T., Heath, D., Roe, I.â•›F., & Götz, D. (2004). A valency dictionary of English. Berlin/New
York: Mouton de Gruyter. DOI: 10.1515/9783110892581
Hilpert, M. (2008). New evidence against the modularity of grammar: Constructions, colloca-
tions, and speech perception. Cognitive Linguistics, 19(3), 483–503.
DOI: 10.1515/COGL.2008.019
Hilpert, M. (2010). The force dynamics of English complement clauses: A usage-based account.
In K. Fischer & D. Glynn (Eds.), Quantitative methods in cognitive semantics: Corpus-driven
approaches (pp. 155–178). Berlin: Mouton de Gruyter.
Hooper, J. (1976). Word frequency in lexical diffusion and the source of morphophonological
change. In W. Christie (Ed.), Current progress in historical linguistics (pp. 96–105). A
� msterdam:
NorthHolland.
Hopper, P. (1987). Emergent grammar. Berkeley Linguistic Society, 13, 139–157.
Hopper, P., & Thompson, S. (1980). Transitivity in grammar and discourse. Language, 56(2),
251–299. DOI: 10.1353/lan.1980.0017
Hunn, E.â•›S. (1975). A measure of the degree of correspondence of folk to scientific biological
classification. American Ethnologist, 2(2), 309–327. DOI: 10.1525/ae.1975.2.2.02a00080
Hunston, S., & Francis, G. (2000). Pattern grammar: A corpus-driven approach to the lexical
grammar of English. Amsterdam: John Benjamins. DOI: 10.1075/scl.4
Iwata, S. (2005). The role of verb meaning in locative alternations. In M. Fried & H. Boas (Eds.),
Grammatical constructions: Back to the roots (pp. 101–118). Amsterdam: John Benjamins.
DOI: 10.1075/cal.4.07iwa
Iwata, S. (2008). Locative alternation: A lexical-constructional approach. Amsterdam: John �Benjamins.
DOI: 10.1075/cal.6
Jackendoff, R. (1972). Semantic interpretation in generative grammar. Cambridge, MA: MIT Press.
Jackendoff, R. (1975). Morphological and semantic regularities in the lexicon. Language, 51(3),
639–671. DOI: 10.2307/412891
Jackendoff, R. (1976). Towards an explanatory semantic representation. Linguistic Inquiry, 7,
89–150.
Jackendoff, R. (1983). Semantics and cognition. Cambridge, MA: MIT Press.
Jackendoff, R. (1990). Semantic structures. Cambridge, MA: MIT Press.
Jackendoff, R. (2002). Foundations of language: Brain, meaning, grammar, evolution. Oxford &
New York: Oxford University Press.
Johnson, K. (2008). Quantitative methods in linguistics. Oxford: Blackwell Publishing.
Jurafsky, D., Bell, A., Gregory, M., & Raymond, W.â•›D. (2001). Probabilistic relations between
words: Evidence from reduction in lexical production. In J. Bybee & P. Hopper (Eds.), Fre-
quency and the emergence of linguistic structure (pp. 229–254). Amsterdam: John Benjamins.
DOI: 10.1075/tsl.45.13jur
Kaschak, M.â•›P. & Glenberg, A.â•›M. (2000). Constructing Meaning: The Role of Affordances and
Grammatical Constructions in Sentence Comprehension. Journal of Memory and Language,
43(3), 508–529.
Kay, P. (2005). Argument-structure constructions and the argument-adjunct distinction. In M. Fried
& H. Boas (Eds.), Grammatical constructions: Back to the roots (pp. 71–98). Â�Amsterdam: John
Benjamins. DOI: 10.1075/cal.4.05kay
Kay, P., & Fillmore, C.â•›J. (1999). Grammatical constructions and linguistic generalizations: The
What’s X doing Y? construction. Language, 75, 1–33. DOI: 10.2307/417472
228 Argument Structure in Usage-Based Construction Grammar
Kilgariff, A. (1997). I don’t believe in word senses. Computers and the Humanities, 31(2), 91–113.
DOI: 10.1023/A:1000583911091
Klotz, M. (2007). Valency rules? The case of verbs with propositional complements. In T. Herbst
& K. Götz-Votteler (Eds.), Valency: Theoretical, descriptive and cognitive issues (pp. 117–128).
Berlin: Mouton de Gruyter.
Koenig, J.-P., Mauner, G., & Bienvenue, B. (2003). Arguments for adjuncts. Cognition, 89(2),
67–103. DOI: 10.1016/S0010-0277(03)00082-9
Kuperman, V., & Bresnan, J. (2012). The effects of construction probability on word durations
during spontaneous incremental sentence production. Journal of Memory and Language,
66, 588–611. DOI: 10.1016/j.jml.2012.04.003
Laffut, A. (1997). The spray/load alternation: Some remarks on a textual and constructionist
approach. Leuvense Bijdragen, 86, 457–487.
Lakoff, G., & Johnson, M. (1980). Metaphors we live by. Chicago: University of Chicago Press.
Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind.
Chicago: University of Chicago Press. DOI: 10.7208/chicago/9780226471013.001.0001
Langacker, R.â•›W. (1987). Foundations of cognitive grammar. Volume 1: Theoretical prerequisites.
Stanford: Stanford University Press.
Langacker, R.â•›W. (1991). Foundations of cognitive grammar. Volume 2: Descriptive application.
Stanford: Stanford University Press.
Langacker, R.â•›W. (2000). A dynamic usage-based model. In M. Barlow & S. Kemmer (Eds.),
Usage-based models of language (pp. 1–63). Stanford: CSLI Publications.
Langacker, R.â•›W. (2008). Cognitive grammar: A basic introduction. Oxford: Oxford University
Press. DOI: 10.1093/acprof:oso/9780195331967.001.0001
Langacker, R.â•›W. (2009). Constructions and constructional meaning. In V. Evans & S. Pourcel
(Eds.), New directions in cognitive linguistics (pp. 225–267). Amsterdam: John Benjamins.
DOI: 10.1075/hcp.24.17lan
Lapata, M., Keller, F., & Schulte im Walde, S. (2001). Verb frame frequency as a predictor of verb
bias. Journal of Psycholinguistic Research, 30(4), 419–435. DOI: 10.1023/A:1010473708413
Lauwers, P., & Willems, D. (2011). Coercion: Definition and challenges, current approaches, and
new trends. Linguistics, 49(6), 1219–1235. DOI: 10.1515/ling.2011.034
Leech, G., Rayson, P., & Wilson, A. (2001). Word frequencies in written and spoken English. London:
Longman.
Lemmens, M. (1998). Lexical perspectives on transitivity and ergativity: Causative constructions
in English. Amsterdam: John Benjamins. DOI: 10.1075/cilt.166
Levin, B. (1993). English verb classes and alternations: A preliminary investigation. Chicago:
University of Chicago Press.
Levin, B., & Rappaport Hovav, M. (2005). Argument realization. Cambridge: Cambridge U � niversity
Press. DOI: 10.1017/CBO9780511610479
Lieven, E.â•›V.â•›M., Pine, J.â•›M., & Baldwin, G. (1997). Lexically-based learning and early grammatical
development. Journal of Child Language, 24, 187–220. DOI: 10.1017/S0305000996002930
Luce, R.â•›D. (1986). Response times. New York: Oxford University Press.
Marcotte, J.-P. (2005). Causative alternation errors in child language acquisition (Doctoral dis-
sertation, Stanford University). Retrieved from http://www.tc.umn.edu/~marco043/files/
MarcotteThesis2005.pdf
Marcotte, J.-P. (2006). Causative alternation errors as event-driven construction paradigm
completions. In E.â•›V. Clark & B. Kelly (Eds.), Constructions in acquisition (pp. 205–232).
Stanford: CSLI Publications.
229
References
Mazurkewich, I., & White, L. (1984). The acquisition of the dative alternation: Unlearning Over-
generalizations. Cognition, 16, 261–283. DOI: 10.1016/0010-0277(84)90030-1
McClure, K, Pine, J.â•›M., & Lieven, E.â•›V.â•›M. (2006). Investigating the abstractness of children’s
early knowledge of argument structure. Journal of Child Language, 33, 693–720.
DOI: 10.1017/S0305000906007525
McRae, K., Ferretti, T., & Amyote, L. (1997). Thematic roles as verb-specific concepts. Language
and Cognitive Processes, 12(2/3), 137–176. DOI: 10.1080/016909697386835
Medin, D.â•›L., & Schaffer, M.â•›M. (1978). Context theory of classification learning. Psychological
Review, 85, 207–238. DOI: 10.1037/0033-295X.85.3.207
Michaelis, L.â•›A., & Ruppenhofer, J. (2001). Beyond alternations: A constructional model of the
German applicative pattern. Stanford: CSLI Publications.
Michaelis, L.â•›A. (2004). Type shifting in construction grammar: An integrated approach to as-
pectual Coercion. Cognitive Linguistics, 15, 1–67. DOI: 10.1515/cogl.2004.001
Michaelis, L.â•›A. (2005). Entity and event coercion in a symbolic theory of syntax. In J.-O. Östman
& M. Fried (Eds.), Construction grammars: Cognitive grounding and theoretical extensions
(pp. 45–87). Amsterdam: John Benjamins. DOI: 10.1075/cal.3.04mic
Miller, G.â•›A. (1995). WordNet: A lexical database for English. Communications of the ACM,
38(11), 39–41. DOI: 10.1145/219717.219748
Müller, S., & Wechsler, S. (2014). Lexical approaches to argument structure. Theoretical Linguis-
tics, 40(1–2), 1–76. DOI: 10.1515/tl-2014-0001
Müller, S. (2006). Phrasal or lexical constructions? Language, 82(4), 850–883.
DOI: 10.1353/lan.2006.0213
Nemoto, N. (2005). Verbal polysemy and frame semantics in construction grammar: some
observations about the locative alternation. In M. Fried & H. Boas (Eds.), Grammatical
constructions: Back to the roots (pp. 119–138). Amsterdam: John Benjamins.
DOI: 10.1075/cal.4.08nem
Nosofsky, R.â•›M. (1986). Attention, similarity, and the identification-categorization relationship.
Journal of Experimental Psychology: General, 115, 39–57. DOI: 10.1037/0096-3445.115.1.39
Nosofsky, R.â•›M. (1988). Similarity, frequency, and category representations. Journal of Experimen-
tal Psychology: Learning, Memory, and Cognition, 14, 54–65. DOI: 10.1037/0278-7393.14.1.54
Nosofsky, R.â•›M., Pothos, E.â•›M., & Wills, A.â•›J. (2011). The generalized context model: An exemplar
model of classification. In E.â•›M. Pothos & A.â•›J. Wills (Eds.), Formal approaches in categoriza-
tion (pp. 18–39). Cambridge: Cambridge University Press.
DOI: 10.1017/CBO9780511921322.002
Ogura, M. (1993). The development of periphrastic do in English: A case of lexical diffusion in
syntax. Diachronica, 10(1), 51–85. DOI: 10.1075/dia.10.1.04ogu
Payne, T.â•›E. (1997). Describing morphosyntax: A guide for field linguists. Cambridge: Cambridge
University Press.
Perek, F., & Goldberg, A.â•›E. (to appear). Generalizing beyond the input: The functions of the
constructions matter.
Perek, F. (2010). Identification de constructions grammaticales en corpus╯ : Une approche quan-
titative de l’augmentation de valence. In P. Cappeau, H. Chuquet, & F. Valetopoulos (Eds.),
Travaux linguistiques du CerLiCO: Vol. 23. L’exemple et le corpus: Quel statut? (pp. 165–180).
Rennes: Presses Universitaires de Rennes.
Perek, F. (2012). Alternation-based generalizations are stored in the mental grammar: Evidence
from a sorting task experiment. Cognitive Linguistics, 23(3), 601–635.
DOI: 10.1515/cog-2012-0018
230 Argument Structure in Usage-Based Construction Grammar
Perek, F. (2014a). Vector spaces for historical linguistics. In Proceedings of the 52nd Annual
Meeting of the Association for Computational Linguistics, Baltimore, Maryland USA, June
23–25 2014.
Perek, F. (2014b). Rethinking constructional polysemy: The case of the English conative con-
struction. In D. Glynn & J. Robinson (Eds.), Polysemy and synonymy. Corpus methods and
applications in cognitive linguistics. Amsterdam: John Benjamins. DOI: 10.1075/hcp.43.03per
Perek, F. (to appear). Using distributional semantics to study syntactic productivity in diachrony:
A case study.
Perek, F., & Lemmens, M. (2010). Getting at the meaning of the English at-construction: The case
of a constructional split. CogniTextes, 5. Retrieved from http://cognitextes.revues.org/331
Pickering, M.â•›J., Traxler, M.â•›J., & Crocker, M.â•›W. (2000). Ambiguity resolution in sentence pro-
cessing: Evidence against frequency-based accounts. Journal of Memory and Language, 43,
447–475. DOI: 10.1006/jmla.2000.2708
Pierrehumbert, J. (2001). Exemplar dynamics: Word frequency, lenition and contrast. In J. Bybee &
P. Hopper (Eds.), Frequency and the emergence of linguistic structure (pp. 137–57). Â�Amsterdam:
John Benjamins. DOI: 10.1075/tsl.45.08pie
Pinker, S., & Prince, A. (1994). Regular and irregular morphology and the psychological status
of rules of grammar. In S.â•›D. Lima, R.â•›L. Corrigan, & G.â•›K. Iverson (Eds.), The reality of
linguistic rules (pp. 353–388). Amsterdam: John Benjamins. DOI: 10.1075/slcs.26.21pin
Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure. Cambridge,
MA: MIT Press/Bradford Books.
Pinker, S. (1991). Rules of language. Science, 253, 530–535. DOI: 10.1126/science.1857983
Pinker, S., Lebeaux, D.â•›S., & Frost, L.â•›A. (1987). Productivity and constraints in the acquisition
of the passive. Cognition, 26, 195–267. DOI: 10.1016/S0010-0277(87)80001-X
Plag, I. (2003). Word-formation in English. Cambridge: Cambridge University Press.
DOI: 10.1017/CBO9780511841323
Pollard, C., & Sag, I.â•›A. (1994). Head-Driven phrase structure grammar. Chicago: University of
Chicago Press.
Ramchand, G.╛C. (2008). Verb meaning and the Lexicon: A first-phase syntax. Cambridge: �Cambridge
University Press. DOI: 10.1017/CBO9780511486319
Randall, J.â•›M. (2010). Linking: The geometry of argument structure. Dordrecht: Springer.
Rappaport Hovav, M., & Levin, B. (1998). Building verb meanings. In M. Butt & W. Geuder
(Eds.), The projection of arguments: Lexical and compositional factors (pp. 97–134). Stanford:
CSLI Publications.
Rappaport Hovav, M., & Levin, B. (2008). The English dative alternation: The case for verb
sensitivity. Journal of Linguistics, 44(1), 129–167.
Rappaport, M., & Levin, B. (1988). What to do with θ-roles? In W. Wilkins (Ed.), Syntax and
semantics: Vol. 21. Thematic relations (pp. 7–36). San Diego: Academic Press.
Ratcliff, R. (1993). Methods for dealing with reaction time outliers. Psychological Bulletin, 114,
510–532. DOI: 10.1037/0033-2909.114.3.510
Reali, F., & Christiansen, M.â•›H. (2007). Processing of relative clauses is made easier by frequency
of occurrence. Journal of Memory and Language, 57, 1–23. DOI: 10.1016/j.jml.2006.08.014
Reddy, M.â•›J. (1979). The conduit metaphor: A case of frame conflict in our language about lan-
guage. In A. Ortony (Ed.), Metaphor and thought (pp. 284–324). Cambridge: Cambridge
University Press.
Rosch, E., Mervis, C.â•›B., Gray, W.â•›D., Johnson, D.â•›M., & Boyes-Braem, P. (1976). Basic objects in nat-
ural categories. Cognitive Psychology, 8(3), 382–439. DOI: 10.1016/0010-0285(76)90013-X
231
References
Salkoff, M. (1983). Bees are swarming in the garden: A systematic synchronic study of productiv-
ity. Language, 59(2), 288–346. DOI: 10.2307/413576
Schlesinger, I.â•›M. (1995). On the semantics of the object. In B. Aarts & C.â•›F. Meyer (Eds.), The verb
in contemporary English: Theory and description (pp. 54–74). Cambridge, MA: Cambridge Â�
University Press.
Schmid, H.-J., & Kűchenhoff, H. (2013). Collostructional analysis and other ways of measur-
ing lexicogrammatical attraction: Theoretical premises, practical problems and cognitive
underpinnings. Cognitive Linguistics, 24(3), 531–577. DOI: 10.1515/cog-2013-0018
Schmid, H.-J. (2010). Does frequency in text instantiate entrenchment? In K. Fischer & D. Glynn
(Eds.), Quantitative methods in cognitive semantics: Corpus-driven approaches (pp. 101–133).
Berlin: Mouton de Gruyter.
Sethuraman, N. (2004). Influence of parental input on learning argument structure construc-
tions. In A. Brugos, L. Micciulla, & C.â•›E. Smith (Eds.), On-line Proceedings supplement
of Boston University Child Development 28. Retrieved from http://www.bu.edu/bucld/
proceedings/supplement/vol28/
Stefanowitsch, A. (2011). Argument structure: Item-based or distributed? Zeitschrift für Anglistik
und Amerikanistik, 59(4), 331–346.
Stefanowitsch, A., & Gries, S.â•›T. (2003). Collostructions: Investigating the interaction of words
and constructions. International Journal of Corpus Linguistics, 8(2), 209–243.
DOI: 10.1075/ijcl.8.2.03ste
Stefanowitsch, A., & Gries, S.â•›T. (2005). Covarying collexemes. Corpus Linguistic and Linguistic
Theory, 1(1), 1–43. DOI: 10.1515/cllt.2005.1.1.1
Stemberger, J.â•›P., & MacWhinney, B. (1986). Frequency and the lexical storage of regularly in-
flected forms. Memory & Cognition, 14(1), 17–26. DOI: 10.3758/BF03209225
Stemberger, J.â•›P., & MacWhinney, B. (1988). Are inflected forms stored in the lexicon? In
M. Hammond & M. Noonan (Eds.), Theoretical morphology: Approaches in modern lin-
guistics (pp. 101–116). San Diego: Academic Press.
Suttle, L., & Goldberg, A. (2011). The partial productivity of constructions as induction. Lin-
guistics, 49(6), 1237–1269. DOI: 10.1515/ling.2011.035
Szmrecsanyi, B. (2006). Morphosyntactic persistence in spoken English. Berlin/New York: Mouton
de Gruyter. DOI: 10.1515/9783110197808
Talmy, L. (1985). Lexicalisation patterns: Semantic structures in lexical forms. In T. Shopen
(Ed.), Language typology and syntactic description. Volume III: Grammatical categories and
the lexicon (pp. 55–149). New York: Cambridge University Press.
Talmy, L. (1996). The windowing of attention in language. In M. Shibatani & S.â•›A. Thompson
(Eds.), Grammatical constructions: Their form and meaning (pp. 235–287). Oxford: Oxford
University Press.
Talmy, L. (2000). Toward a cognitive semantics. Cambridge, MA: MIT Press.
Tao, H. (2001). Discovering the usual with corpora: The case of remember. In R. Simpson &
J. Swales (Eds.), Corpus linguistics in North America: Selections from the 1999 symposium
(pp. 116–144). Ann Arbor: University of Michigan Press.
Tao, H. (2003). A usage-based approach to argument structure: ‘Remember’ and ‘forget’ in
spoken English. International Journal of Corpus Linguistics, 8(1), 75–95.
DOI: 10.1075/ijcl.8.1.04tao
Taylor, J.â•›R. (1995). Linguistic categorization. Oxford: Oxford University Press.
Tenny, C. (1994). Aspectual roles and the syntax-semantics interface. Dordrecht: Kluwer.
DOI: 10.1007/978-94-011-1150-8
232 Argument Structure in Usage-Based Construction Grammar
Wilson, M.â•›P., & Garnsey, S.â•›M. (2009). Making simple sentences hard: Verb bias effects in simple
direct object sentences. Journal of Memory and Language, 60, 368–392.
DOI: 10.1016/j.jml.2008.09.005
Wonnacott, E. (2011). Balancing generalization and lexical conservatism: An artificial language
study with child learners. Journal of Memory and Language, 65, 1–14.
DOI: 10.1016/j.jml.2011.03.001
Wonnacott, E., Boyd, J., Thompson, J., & Goldberg, A.â•›E. (2012). Input effects on the acquisi-
tion of a novel phrasal construction in 5 year olds. Journal of Memory and Language, 66,
458–478. DOI: 10.1016/j.jml.2011.11.004
Wonnacott, E., Newport, E.â•›L., & Tanenhaus, M.â•›K. (2008). Acquiring and processing verb argu-
ment structure: Distributional learning in a miniature language. Cognitive Psychology, 56,
165–209. DOI: 10.1016/j.cogpsych.2007.04.002
Zeldes, A. (2013). Productive argument selection: Is lexical semantics enough? Corpus Linguistics
and Linguistic Theory, 9(2), 263–291. DOI: 10.1515/cllt-2013-0006
Zeschel, A. (2009). What’s (in) a construction? In V. Evans & S. Pourcel (Eds.), New directions
in Cognitive Linguistics (pp. 185–200). Amsterdam: John Benjamins.
DOI: 10.1075/hcp.24.15zes
Zeschel, A. (2012). Incipient productivity. A construction-based approach to linguistic creativity.
Berlin/New York: de Gruyter. DOI: 10.1515/9783110274844
Zwicky, A.â•›M. (1971). In a manner of speaking. Linguistic Inquiry, 2(2), 223–233.
Appendix
Stimuli used in the productivity
experiment (Chapter 7)
I polysemyâ•… 16, 21, 27, 50, 86, 98, semantic roleâ•… 16, 20
inheritanceâ•… 153, 155 100, 150 semantic structureâ•… 18–21, 169
introspectionâ•…216–217 constructional polysemy statistical preemptionâ•… 147f,
111–112 149f, 152, 195–196, 207–208
L primingâ•… 165–167, 171, 187 subcategorization frameâ•… 16
language acquisitionâ•… 17, 75, 81, constructional primingâ•… 166 surface generalizations
109, 152, 168 productivityâ•… 7, 9, 95, 168, hypothesisâ•…148–149
lexical entryâ•… 16, 21, 23, 29–34, 175–176, 198, 205 syntactic frameâ•… 18, 23, 31, 33
38–40 alternation-based
lexical ruleâ•… 23, 148, 169 productivityâ•…169–170, V
lexical unitâ•… 16, 119, 150 172, 203 valencyâ•… 1f, 16, 29, 31, 45–46
linking ruleâ•… 18–21 argument structure valency constructionâ•… 36–37
productivityâ•… 76, 148, valency grammarâ•… 15, 17, 35
M 168–170 valency patternâ•… 35–37,
maze taskâ•… 53–54 asymmetryâ•… 171, 189, 193 45–46, 51–52, 65, 69–73
metaphorâ•… 25, 120–121, 155 construction-based valency slotâ•… 35–36
see also metaphorical productivityâ•… (verb) constantâ•… 21
extension under extension 169–170, 172 verb-specific constructionâ•…
mini-constructionâ•…34 profilingâ•… 30–32, 50, 55, 212 76, 109
multiple argument
realizationâ•… 20–21, 27, 148 R W
rule/list fallacyâ•… 6, 108 windowing of attentionâ•… 28–29,
O 31, 45
overgeneralizationâ•… 2, 17, 109, S WordNetâ•…118–121
169–170 scene encoding hypothesisâ•… 98
schematizationâ•… 6, 105, 115, 168
P semantic compatibilityâ•… 24–25,
parsimonyâ•… 33, 108 85–86, 140, 168–170
participant roleâ•… 24–29, 36, 41, semantic frameâ•… 16, 24, 28–33,
45, 68–69 41, 45, 118–119, 150, 217