Anda di halaman 1dari 10

19

Representing the Adverb Very in Fuzzy


Set Theory
Martine De Cock
Applied Mathematics and Computer Science, University of Gent
Krijgslaan 281 (S9), B-9000 Gent, Belgium
Martine.DeCock@rug.ac.be
Abstract. We recall the concept of a linguistic variable and the representation of
its values (i.e. linguistic terms) by means of fuzzy sets. In this framework adverbs
are represented by fuzzy modiers, i.e. operators acting on these fuzzy sets. We
investigate two important classes of popular fuzzy modiers. Grounding on results
from psycholinguistic research, we discuss their pros and contras for representing
the adverb very.
19.1 Linguistic variables
19.1.1 The concept of a linguistic variable
The concept of a linguistic variable was introduced by Zadeh in the 70s
[ZAD1975]. While the values of a numerical variable (often used in classical
mathematics, physics, economics, ...) are numbers, the values of a linguistic
variable are linguistic terms. E.g. the numerical variable size has values
5 cm
2
, 100 cm
2
, ... while the linguistic variable Size has values large,
small, very large, rather small, not very large and not small, .... The set
consisting of all possible values of a linguistic variable is called the term
set. Perhaps the most important computational beauty of the concept of a
linguistic variable is the fact that the elements of its term set show a specic
structure. Starting from one, two, or more base terms (e.g. large, small),
every other term can be constructed using the following scheme :
223
Proceedings of the ESSLLI Student Session 1999
Amalia Todirascu (editor).
Chapter 19, Copyright c 1999, Martine De Cock.
Representing the Adverb Very in Fuzzy Set Theory: Martine De Cock /224
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
large
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
close to 5
Figure 19.1: Fuzzy sets for large and close to 5
<conjunction> := and;
<disjunction> := or;
<modier> := very | rather | more or less | extremely |
quite | sort of | really | slightly | ...;
<term> := <base term> | <term> <conjunction> <term> |
<term> <disjunction> <term> |
not <term> | <modier> <term>;
19.1.2 Representing the meaning of a term by a fuzzy set
The meaning of each term can be represented by a fuzzy set. If
X is a universe of discourse, then a fuzzy set A in X is characterized by its
membership function
A : X [0, 1]
x A(x), x X
The membership function of A maps every object x of the universe X onto
a degree of membership, i.e. the degree to which x belongs to the fuzzy
set A. F(X) is the set of all fuzzy sets in X. E.g. if X = [0, 12] then the
terms large and close to 5 could be represented by the fuzzy sets in g. 19.1.
For simplicity we will make no notational distinction between a fuzzy set A,
its membership function A and the term A represented by that fuzzy set.
Once the membership function for the base terms is known, the membership
function of all other terms can be deduced.
Representation of conjunction and disjunction The conjunctor and
and the disjunctor or can be modelled using min and max respectively. If A
and B are two fuzzy sets representing the meaning of two linguistic terms
225\ ESSLLI Student Session 1999
then the fuzzy sets for A and B and for A orB can be derived in the
following way :
(x X)((A and B)(x) = (A
min
B)(x) = min(A(x), B(x)))
(x X)((A or B)(x) = (A
max
B)(x) = max(A(x), B(x)))
Representation of linguistic modiers In this paper however we will
focus on the deduction of the membership function of a modied term from
the membership function of that term. With a linguistic modier, i.e. an
adverb in natural language, a fuzzy modier m is associated, i.e. a F(X)
F(X) mapping
m : F(X) F(X)
A m(A), A F(X)
So a fuzzy modier is an operator acting on a fuzzy set, transforming a
fuzzy set into another one in the same universe.
Inclusion can be dened for fuzzy sets A F(X), B F(X) in the
following way :
A B (x X)(A(x) B(x))
So the fuzzy set A is a subset of the fuzzy set B i for every object x the
degree to which x belongs to A doesnot exceed the degree to which x belongs
to B. Now we can distinguish two important subclasses of fuzzy modiers :
m is expansive (A F(X))(A m(A))
m is restrictive (A F(X))(m(A) A)
A restrictive modier leads to a decrease in the degrees of membership; it
pushes the original membership function down. Bouchon and Jia [BOU1992]
call this reinforcement, Lako[LAK1973] talks about intensiers. Re-
strictive modiers are thus often associated with intensifying adverbs such
as very, extremely, highly, ... An expansive modier on the other hand leads
to an increase of the degrees of membership; it lifts the original mem-
bership function up. Bouchon and Jia[BOU1992] call this weakening,
Lako[LAK1973] talks about deintensiers. Expansive modiers are thus
often associated with deintensifying adverbs such as more or less, rather, a
bit, ... In this paper we will focus on the representation of one particular ad-
verb, namely very. We will among other things explain that it cant always
be represented by a restrictive fuzzy modier.
Representing the Adverb Very in Fuzzy Set Theory: Martine De Cock /226
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
large
very large
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
close to 5
very close to 5
Figure 19.2: Eect of a powering modier on large and close to 5
19.2 Two popular classes of fuzzy modiers
19.2.1 Powering modiers
In the early 70s Zadeh[ZAD1972] introduced a class of fuzzy powering mod-
iers that has become very popular. For [0, +[, P

is a F(X) F(X)
mapping with for A F(X)
P

(A) : X [0, 1]
x (A(x))

, x X
If 1 then P

is a restrictive modier, i.e. (A F(X))(P

(A) A).
Nowadays the fuzzy modier P
2
is still a very popular choice for representing
the adverb very (see e.g. [BAB1998], [YASMUK1998]). Fig. 19.2 shows the
eect of P
2
on the fuzzy sets large and close to 5.
19.2.2 Shifting modiers
Another type of fuzzy modiers, the shifting modiers, was already infor-
mally suggested by Lako[LAK1973] in the 70s. A.o. Hellendoorn [HEL1990]
and Bouchon[BOU1993] used it in a more formal manner. Since the shifting
is an operation on objects of the universe (and not an operation on their
degree of membership like the powering is), its only applicable for fuzzy
sets in a universe equipped with such an operation. For X = R, i.e. the
set of real numbers, and R we dene the shifting modier S

with for
A F(R)
S

(A) : R [0, 1]
x A(x ), x R
227\ ESSLLI Student Session 1999
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
large
very large
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12
close to 5
A
B
Figure 19.3: Eect of shifting modiers on (a) large and (b) close to 5
19.3 Shortcomings of the fuzzy modiers as a rep-
resentation for very
19.3.1 Behaviour of S

with regard to restrictiveness


Unlike with the powering modiers, we cant distinguish a set of values for
for which S

is restrictive. Let A F(R) then the following properties


hold
1
:
If A is increasing and 0 then S

(A) A
If A is decreasing and 0 then S

(A) A
If A is increasing, then very A could be represented by e.g. S
1
. If A is
decreasing, then very A could be represented by e.g. S
1
. But in general,
there is no so that (A F(X)) (S

(A) A). Fig. 19.3a shows how


we can generate the membership function for very large from the increasing
membership function for large using S
1
. The membership function for close
to 5 however is neither increasing or decreasing; we cannot use a shifting
modier S

to generate very close to 5. Fig. 19.3b illustrates the eect of


S
1
and S
1
on close to 5 (A = S
1
(close to 5), B = S
1
(close to 5)).
19.3.2 Behaviour of P
2
with respect to kernel and support
For A F(X) :
kernel A = {x|x X A(x) = 1}
support A = {x|x X A(x) > 0}
1
We recall that for A F(R)
A is increasing ((x, y) R
2
)(x y A(x) A(y))
A is decreasing ((x, y) R
2
)(x y A(x) A(y))
Representing the Adverb Very in Fuzzy Set Theory: Martine De Cock /228
Probably the best-known shortcoming of the powering modiers is the fact
that they keep the kernel and the support ([KER1993] p. 35, [HEL1990]
p. 38, [LAK1973] p. 488, [DES1988] p. 64). For [0, +[ :
(A F(X))(kernel(P

(A)) = kernel A support(P

(A)) = support A)
If we choose to represent very by P
2
, then every square x that is certainly
large (x kernel(large)) is also very large (x kernel(very large)). In other
words in this representation there cant be any squares that are considered
to be large to degree 1, but very large to a lower degree. Every square
that is large to degree 1, is automatically considered to be very large to
degree 1 as well. In some reallife situations however a square with a size
of 7.5cm e.g. can be considered large to degree 1, but not very large to
the same degree. (Compare with : some people consider a man of 75 years
certainly old (75 kernel(old)), but not certainly very old (75 / kernel(very
old).) Furthermore : every square x that isnt very large to any degree (x /
support(very large)), cant be considered to be large in any degree either
(x / support(large)). These properties are clearly counterintuitive. Its
worth mentioning that the shifting modiers do not have this shortcoming.
19.3.3 Two interpretations of very
In the mid 70s, Hersh and Caramazza[HECA1976] did some psycholinguistic
experiments on the accuracy of P
2
as a representation for very (and the
usefulness of fuzzy sets to model linguistic terms in general). One of their
most interesting ndings was that there are two dierent interpretations of
the adverb very. Vanden Eynde[VDEYN] also did research on this subject
and came to a similar conclusion. The two interpretations are
1. The inclusive interpretation (g. 19.4a) : the fuzzy set very large
is included in the fuzzy set large. Semantical entailment is clearly
respected : (x X)(x is very large x is large).
2. The noninclusive interpretation (g. 19.4b) : the fuzzy set very
large isnt included in large, neither is large in very large. large and
very large denote two dierent (overlapping) categories. This doesnt
mean that the person interpreting large and very large doesnt respect
semantical entailment, but in this case he is following Grices maxime
of conversation [GRI1978] : Make your speech contribution as infor-
mative as required. When a listener hears that x is large, he assumes
that x isnt very large, because in the latter case the speaker would
have used the more informative utterance x is very large. So large
+ > not very large.
We would like to remark that the dierence between the two interpreta-
tions of large and very large doesnt only have consequences for the choice
229\ ESSLLI Student Session 1999
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
large
very large
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12
large
very large
Figure 19.4: (a) Inclusive interpretation (b) noninclusive interpretation
of a fuzzy modier to model very (restrictive or not), but also inuences the
shape of the membership function for large.
It is clear that Zadehs restrictive P
2
can only be used to model very in
the inclusive interpretation. Finally we would like to mention that Hersh
and Caramazzas research showed another shortcoming of Zadehs P
2
in the
inclusive interpretation. While P
2
increases the slope of an increasing mem-
bership function, the slopes of the functions for tall and very tall resulting
from the experiment were approximately equal. In fact Hersh and Cara-
mazza decided that the membership function for tall could be shifted to the
right to obtain a pretty good approximation for the membership function
for very tall.
19.3.4 Choosing the best representation
The two kinds of fuzzy modiers described above, i.e. the powering and the
shifting modiers, are easy operators from a mathematical point of view.
However they both have their own advantages and disadvantages from a
linguistic perspective. Neither one of them is the best in general, but we
can give some guidelines on how to choose the most suitable in a specic
situation. We assume that in a noninclusive interpretation the fuzzy sets
have membership functions that can be approximated by a Gaussian curve
like in g. 19.4b, while in an inclusive interpretation they can have any shape.
Furthermore we keep in mind that P
2
has the disadvantage of keeping the
kernel and the support, so well avoid the use of P
2
whenever possible. For
X a universe and A F(X) we can use the following guidelines to decide
which fuzzy modier we will use to represent very :
Representing the Adverb Very in Fuzzy Set Theory: Martine De Cock /230
start

no yes
inclusive interpretation? X equipped with shifting? S

yes no
no
X equipped with shifting? P
2
?
yes
no
A increasing or decreasing? P
2
yes
S

19.3.5 Conclusion and future work


The adverb very can be modelled in a lot of situations using powering or
shifting modiers. The choice between those two depends mainly on the in-
terpretation (inclusive or noninclusive) and the kind of universe (equipped
with a shifting operation or not). Some cases however can not be handled
(cfr. a noninclusive interpretation in a universe not equipped with a shifting
operation). Furthermore, even in situations where very can be represented
using P
2
, this fuzzy modier has the great disadvantage of keeping the kernel
and the support.
These observations encourage us to search for another kind of fuzzy
modier. Among other things we would like that this new type of fuzzy
modier :
Doesnt keep the kernel and the support in every case.
Is applicable in all kinds of universes (not only in those equipped with
a special operation, like shifting).
Provides a framework in which both the inclusive and the noninclusive
interpretation can be modelled in a similar mathematical way.
Acknowledgement
The author would like to thank the Fund for Scientic Research Flanders
(FWO) for funding the research reported on in this paper.
Bibliography
[BAB1998] R. BABU

SKA, Fuzzy Modeling for Control, International


Series in Intelligent Technologies, Kluwer Academic Publishers, (1998)
[BOU1992] B. BOUCHON-MEUNIER, Y. JIA, Linguistic Modiers and
Imprecise Categories, International Journal of Intelligent Systems,
7(1992), 25-36
[BOU1993] B. BOUCHON-MEUNIER, La Logique Floue, Que sais-je?,
2702, Paris, (1993)
[DES1988] S. DESPR

ES, Un apport `a la conception de syst`emes ` a


base de connaissances : les operations de deduction oues,
Th`ese de doctorat de lUniversite Pierre et Marie Curie (Paris VI),
(1988)
[GRI1978] H. P. GRICE, Further notes on logic and conversation,
P. Cole (ed.), Syntax and Semantics 9 : Pragmatics, New York : Aca-
demic Press 1978, 113-128
[HEL1990] H. HELLENDOORN, Reasoning with fuzzy logic,
Ph. D. thesis, T.U. Delft, 1990
[HECA1976] H.M. HERSH, A. CARAMAZZA, A Fuzzy Set Approach
to Modiers and Vagueness in Natural Language, Journal of
Experimental Psychology : General, 105(1976),nr. 3, 254-276
[KER1993] E. E. KERRE, Introduction to the Basic Principles of
Fuzzy Set Theory and Some of its Applications, Communication
and Cognition, Gent, (1993)
[LAK1973] G. LAKOFF, Hedges : a Study in Meaning Criteria
and the Logic of Fuzzy Concepts, Journal of Philosophical Logic,
2(1973), 458-508
[VDEYN] C. VANDEN EYNDE, A very dicult problem : modelling
modication using VERY. A semanticpragmatic approach.,
(in Dutch) FKFOproject, private communication
231
Proceedings of the ESSLLI Student Session 1999
Amalia Todirascu (editor).
Chapter 19, Copyright c 1999, Martine De Cock.
Representing the Adverb Very in Fuzzy Set Theory: Martine De Cock /232
[YASMUK1998] H. YASUI, M. MUKAIDONO, Fuzzy Prolog based on
Lukasiewicz implication, Logic Programming and Soft Computing,
edited by T.P. Martin, F. Arcelli Fontana; Research Studies Press Ltd.,
(1998)
[ZAD1972] L. A. ZADEH, A Fuzzy-Set-Theoretic Interpretation of
Linguistic Hedges, Journal of Cybernetics, 2,3(1972), 4-34
[ZAD1975] L. A. ZADEH, The Concept of a Linguistic Variable and
its Application to Approximate Reasoning I, II, III, Information
Sciences, 8(1975), 199-249, 301-357, 9(1975), 43-80

Anda mungkin juga menyukai