Anda di halaman 1dari 22

Int. J. Human-Computer Studies (2000) 52, 1 } 22 Article No. ijhc.1999.0325 Available online at http://www.idealibrary.

com on

The impact of animated interface agents: a review of empirical research


DORIS M. DEHN* Department of Psychology, University of the Saarland, P.O. Box 151150, D-66041 Saarbru K cken, Germany. email: doris@cops.uni-sb.de SUSANNE VAN MULKEN German Research Center for Artixcial Intelligence (DFKI), Stuhlsatzenhausweg 3, D-66123 Saarbru K cken, Germany. email: mulken@dfki.de (Received 24 August 1998 and accepted in revised form 1 June 1999)
Over the last years, the animation of interface agents has been the target of increasing interest. Largely, this increase in attention is fuelled by speculated e!ects on human motivation and cognition. However, empirical investigations on the e!ect of animated agents are still small in number and di!er with regard to the measured e!ects. Our aim is two-fold. First, we provide a comprehensive and systematic overview of the empirical studies conducted so far in order to investigate e!ects of animated agents on the user's experience, behaviour and performance. Second, by discussing both implications and limitations of the existing studies, we identify some general requirements and suggestions for future studies. 2000 Academic Press

1. Introduction
Current research in software technology shows a development toward user interfaces in which communication between human and computer is partly mediated by interface agents (e.g. Maes, 1994). Interface agents are computer programs that aid a user in accomplishing tasks carried out at the computer, such as sorting email, "ltering information and scheduling meetings. These agents di!er from conventional computer programs in that they can act autonomously on behalf of the user, that is, without requiring the user to enter a command or click a button whenever she wants the task to be carried out. In addition to autonomy, a characteristic of intelligent agents is their ability to perform tasks delegated to them in an intelligent, that is, context- and user-dependent way. Advances in computer graphics have made feasible the realization of animations based on either real video, cartoon-style drawings or model-based three-dimensional (3D) graphics (e.g. Trappl & Petta, 1997). As a consequence, a more speci"c line of research has recently started to focus on animated interface agents. Such agents appear on the

* Current address: Institute for Aeronautics and Astronautics, Berlin University of Technology, Marchstr. 12, 10587 Berlin, Germany.

1071-5819/00/010001#22 $35.00/0

2000 Academic Press

D. M. DEHN AND S. VAN MULKEN

screen as embodied characters and exhibit various types of life-like behaviours, such as speech, emotions, gestures and eye, head and body movements. The growing interest in animated agents is indicated by workshops organized around this topic (e.g. Andre H & Rist, 1997; Sullivan & Cassell, 1998), dedicated special issues of journals (e.g. Andre H , 1999) and research projects de"ned to investigate their possibilities. Besides the theoretical interest in animated agents, there are a number of systems in which animated agents are already employed. Applications can be found in the form of desktop assistants that aid users in managing desktop tasks*e.g. Peedy the Parrot (Ball et al., 1997)*, presentation agents that present information collected from the world wide web*e.g. PPP-Persona (Rist, Andre H & Mu K ller, 1997)*, pedagogical assistants that guide students through educational or information programs*e.g. Herman the Bug (Lester & Stone, 1997); Gandalf (Cassell et al., 1994); and Olga (Beskow & McGlashan, 1997)*, agent systems that help to facilitate children's mastery of character*e.g. Improv Puppets (Huard & Hayes-Roth, 1997)*and assistants in virtual reality environments*e.g. Steve (Rickel & Johnson, 1999). However, there still is some controversy in the AI community as regards the use of animated agents in user interfaces (cf. Laurel, 1990; Shneiderman & Maes, 1997; Wilson, 1997). Advocates of animated agents appear to be quite optimistic as to the possibilities o!ered by this new style of interaction.
&&Animated pedagogical agents o!er great promise for knowledge-based learning environments. In addition to coupling feedback capabilities with a strong visual presence, these agents play a critical role in motivating students. The extent to which they exhibit life-like behaviours strongly increases their motivational impact 2'' (Lester & Stone, 1997, p. 16). &&Motivation is a key ingredient in learning, and emotions play an important role in motivation. Therefore, we believe that pedagogical agents will be more e!ective teachers if they display and understand emotions.'' (Elliot, Rickel & Lester, 1997, p. 113) &&We surmise that once people are accustomed to synthesized faces, performance becomes more e$cient, and a long partnership further improves performance. Human-like characterization is one good form of autonomous agents, because people are accustomed to interact with other humans.'' (Takeuchi & Naito, 1995, p. 454)

Broadly speaking, advocates of animated agents assume that such agents render a computer system more human-like, engaging, and motivating. By rendering the system more human-like, users can rely on standard interaction skills (such as interpreting their partner's facial expressions or taking into account eye contact) which makes the interaction with the computer much smoother. Coupled with an enhanced motivation to interact with the system (due to a higher degree of entertainingness), this should also support cognitive functions such as problem solving, understanding and learning. In contrast, opponents of animated agents argue that humanizing interface agents might hamper human-computer interaction. One argument capitalizes on the fact that humanization of the interface may induce false mental models of the system (Norman, 1994; Shneiderman & Maes, 1997; Wilson, 1997). For instance, the human-like behaviour of the agent in some aspects may lead the user to believe that the agent resembles human beings in other cognitive and emotional aspects as well. Because of this generalization, the user may ascribe capacities to the system that it does not possess, thus leading to wrong expectations about the system's behaviour. Further problems arise if animations used to render an agent more lively do not map onto system behaviour. For

THE IMPACT OF ANIMATED INTERFACE AGENTS

instance, there are systems in which inactivity is visualized by a moving agent, that is, an agent that taps its foot or makes other types of the so-called idle-time movements. This, however, may lead the user to mistake the agent's activity for system activity (cf. Wilson, 1997). As a consequence of such misinterpretations by the user, the interaction with the system might become less e$cient. A further argument put forward by opponents of animated agents focuses on required cognitive resources: having a further (eye-catching) object to attend to in the interface might merely constitute another potential source of distraction (cf. Walker, Sproull & Subramani, 1994). In addition, there is some empirical evidence suggesting that animated graphics may have a detrimental e!ect on user performance, for instance, in that they hamper text retention (Wright, Milroy & Lickorish, 1999). A standpoint less concerned with possible negative e!ects on the e$ciency of human-computer interaction is that there is in fact no need to represent interface agents as embodied characters. Users tend to treat the computer as a social agent even if it is not visually represented as a human being (Nass, Steuer & Tauber, 1994; Nass, Moon, Fogg, Reeves & Dryer, 1995). Both views on the usefulness of animated agents are potentially testable by empirical means. There are currently several studies that aim at empirically testing the claims put forward by advocates of animated agents. However, it appears to be di$cult to relate the results of the existing studies to each other. One reason for this di$culty is that some of the existing studies yielded contrasting results. For instance, in one study, a system with an animated agent was rated more positively on likeability compared to a system without an animated agent (Koda & Maes, 1996), whereas in another study it was rated less positively (Sproull, Subramani, Kiesler, Walker & Waters, 1996). Another reason is that the empirical studies di!er with regard to the experimental approach chosen to investigate the e!ects of animated agents in human}computer interaction. The di!erence in experimental approaches concerns the following aspects.

he kind of animation used. For instance, in some studies, 2D caricature drawings are employed, whereas in others, realistic 3D models of faces are employed. This aspect concerns the realization of the experimental treatment. he kind of comparison made. For instance, a system with an animated agent can be compared with the same system without any visualization or one that uses graphical symbols instead of the animated agent. This aspect relates to the choice of the control condition. he measures taken as indicators of an e+ect. E!ects can be measured in terms of preference ratings, or they can be measured in terms of di!erences in the user's performance. This aspect concerns the choice of the dependent variable. he task to be carried out in the interaction with the system. For instance, the user might be instructed to achieve some clearly de"ned goal (such as to play poker against the system, cf. Koda & Maes, 1996) or he or she might just be instructed to attend to the information presented by the system (cf. van Mulken, Andre H & Mu K ller, 1998). This aspect relates to the choice of the experimental setting. All these di!erences render it di$cult to obtain a consistent notion of the role of animated agents in human}computer interaction. In this paper, we aim at clarifying the current situation by presenting a structured overview of the empirical studies so far

D. M. DEHN AND S. VAN MULKEN

which allows us to integrate the existing evidence. The classi"cation we use is based on the measured e!ects (i.e. the dependent variables chosen). We also try to point out di!erences in the experimental procedure, especially in those cases in which the empirical evidence is contradictory. In doing so, we hope to resolve inconsistencies in the conclusions drawn by di!erent authors, identify preconditions for e!ects of animated agents to become manifest and formulate some general conclusions on the research on animated agents. The structure of the paper is as follows: We "rst summarize the empirical studies conducted to evaluate the e!ect of animated agents. We then try to answer the question whether the use of animated agents in the interface provides any advantages over not using them. Finally, we formulate several general requirements for future research that might serve to obtain deeper insight in the e!ects of animated agents in human-computer interaction.

2. Empirical studies evaluating the effect of animated agents


The general and rather vague idea that the use of animated agents might improve human-computer interaction can be speci"ed by classifying the various possible e!ects on the user.

One could assume that the use of an animated agent (in comparison to not using one) has a positive e!ect on the user1s subjective experience of the system. For instance, a system with an animated agent could be perceived as more interesting, entertaining or informative than a system without an animated agent. The use of an animated agent could change the user1s behaviour while interacting with the system in some desirable way. For instance, it might be possible that a user spends more time with a system incorporating an agent than with a system without one. The use of an animated agent might improve the outcome of the interaction as indicated by performance data. For instance, a tutoring system that uses an animated agent might result in a better understanding or a more extensive knowledge of the domain in question than a system without an agent.

Although all of the above e!ects are potentially related, they refer to di!erent variables and, hence, require di!erent methods for measuring them.
2.1. THE USER'S SUBJECTIVE EXPERIENCE OF THE SYSTEM

Using an animated agent in a computer system potentially in#uences the user's attitudes towards the system or towards the task which it is supposed to support. A straightforward way of measuring the user's attitude towards the system or the task is to provide him or her with a questionnaire that contains questions about certain characteristics of the system and to collect the corresponding responses. These responses are usually measured as ratings on a 5- or 7-point scale ranging from &&characteristic is fully present'' to &&characteristic is fully absent''. In fact, the use of questionnaires appears to be the predominant approach to quantifying the e!ects of anthropomorphization on the user (e.g. Koda & Maes, 1996; Lester, Converse, Kahler, Barlow, Stone & Bhogal, 1997a; Cassell & Tho H risson, 1999). The dimensions addressed in those questionnaires

THE IMPACT OF ANIMATED INTERFACE AGENTS

commonly pertain to the perceived intelligence, likeability and activity of the system or agent, to the entertainingness and usefulness of the system and to the degree of comfortability during interaction. In the following, we describe the empirical "ndings for these and other dimensions in greater detail. 2.1.1. Intelligence King and Ohya (1996) showed that human forms presented on a computer screen are assessed as more intelligent than simple geometric forms. In this study, however, subjects did not interact with the system and thus did not have any information on particular objects except for their physical appearance. Under this circumstance, it appears almost trivial that*corresponding to the subjects' pre-existing world knowledge*a geometrical object is rated as less intelligent than an anthropomorphized shape. If subjects have the opportunity to base their judgement on other indicators (such as the quality of the system's responses), the visualization of the interface becomes less important. Evidence for this assumption was presented by Koda and Maes (1996). The authors showed that di!erent sorts of visualizations result in di!erent degrees of perceived intelligence, if subjects are merely presented with these visualizations. For instance, a caricature dog was found to be rated less intelligent than a caricature male. Nevertheless, if the subjects played poker against a virtual opponent before rating its intelligence there was no e!ect of the visualization any more. In this situation, an opponent who was not visualized and one who was visualized with a caricature face were assigned the same degree of intelligence. The latter "nding corresponds to the "ndings of Sproull et al. (1996). Sproull et al. had their subjects interact with a career-counselling system, before asking them to rate the virtual counsellor on some personality dimensions. The system's reactions were either presented just as written text on the computer screen or additionally spoken by an animated face on the screen, with the face having either a stern or a pleasant expression. Sproull et al. found no di!erences in the perceived intelligence of the system as rated by their subjects. 2.1.2. Believability A dimension closely related to*though not identical with*intelligence is the perceived believability of the information provided by the system. It can be assumed that a subject is likely to attribute a high degree of believability to a system if it is perceived as intelligent and competent. Lester et al. (1997a) found that the believability of an agent is numerically rated higher if it is fully expressive (i.e. if it conveys di!erent types of feedback both verbally and by means of animation) than if it is muted. The same holds for the perceived knowledge of both agents. Nevertheless, because of a confounding of the system's degree of animation and the amount of advice the agent provided, the di!erence in the ratings cannot be attributed uniquely to the degree of animation. 2.1.3. Likeability/social evaluation Sproull et al. (1996) found that in their career-counselling system, the virtual counsellor was rated higher in social evaluation (e.g. attractiveness and friendliness) and sociability if the subjects were only presented with a textual output rather than additionally with a (stern or pleasant) face. As opposed to this, Koda and Maes (1996) report higher likeability ratings for a poker player visualized by a face than for an invisible poker player. A possible explanation for the inconsistent results might consist in the particular

D. M. DEHN AND S. VAN MULKEN

anthropomorphization chosen: it cannot be ruled out, that the di!erence in likeability ratings for the 2D caricature male (Koda & Maes) or the realistic animated 3D face (Sproull et al.) in comparison to a non-visualized agent are due to the speci"c faces chosen (i.e. to material e!ects) rather than pointing to a general likeability di!erence for visualized vs. non-visualized agents. It has to be acknowledged, though, that Sproull et al. tried to limit the possibility of material e!ects by using two faces with di!erent expressions. Further, in contrast to the Koda and Maes study, subjects in the study by Sproull et al. could have based their judgement also on the voice with which the output was given rather than on the face. Due to its lack of in#ection (cf. Sproull et al., 1996, p. 117), this voice may have sounded odd. 2.1.4. Activity/agency As demonstrated by King and Ohya (1996), agency is signi"cantly more often attributed to anthropomorphized "gures than to simple geometric "gures. Sproull et al., on the other hand, found that a virtual counsellor who is only represented by textual output is rated as signi"cantly more active than the one who is additionally represented by a face and spoken language. Provided that agency and activity tap on similar underlying concepts, the two results appear inconsistent. Again, this inconsistency might be due to the fact that, in the King and Ohya study, subjects could only base their rating on physical properties of the stimulus. In the Sproull et al. study, though, subjects might also base their judgement on the system's behaviour as a career counsellor. Further, as with the perceived likeability, the low rating in activity might be due to the particular face and (slow and unin#ected) voice used in the Sproull et al. study. 2.1.5. Engagingness/entertainment value The most intuitive and compelling assumption is that an animated agent enhances the entertainment value in human}computer interaction. Consistent with this assumption, Takeuchi and Naito (1995) showed that a virtual card matching game with an animated face is (numerically) perceived as more entertaining than a system which uses an arrow to visualize the opponent's moves. Similarly, Koda and Maes found that a poker game with an opponent visualized by a face is rated as more engaging than one without this visualization. There is evidence, though, that the e!ect of an anthropomorphized agent on the entertainment value is domain-speci"c. In a study by van Mulken et al. (1998), subjects experienced a system with an animated agent (the PPP persona) as more entertaining than a system with a pointing arrow if the system served to explain a technical device (a pulley system). If, however, the system had the function of introducing "ctitious employees of a research institute, the subjects' entertainment ratings of the two interfaces did not show any substantial di!erence. In the latter task, subjects were presented with photos of "ctitious sta! members. If we assume that the display of a face raises the entertainment value, then the interface in the employee task might have been entertaining from the start. In this case, an additional anthropomorphized agent might not make a di!erence to the user. This explanation attributes the entertainment advantage to the use of a face rather than the use of animation. Focusing on the e!ect of animation, Lester et al. (1997a) found no statistically di!erence in entertainment ratings for a system with an agent that gives fully animated advice and those for a system with a muted agent.

THE IMPACT OF ANIMATED INTERFACE AGENTS

2.1.6. Comfortability Sproull et al. (1996) found that users of the career-counselling system reported themselves to be more relaxed and self-assured in the interaction if the system's responses were displayed as text than if the responses were uttered by an animated 3D face. This "nding is consistent with "ndings of real social interaction: here, the presence of another person usually enhances arousal in an individual who is asked to perform a task (cf. Zajonc, 1965). In contrast to this, Koda and Maes (1996) report signi"cantly higher comfortability ratings for an animated caricature 2D face as compared to an invisible poker playing opponent. The reason for this inconsistent "nding may lie in the possibility that the caricature face in the Koda and Maes study was not perceived as a real social agent, and so did not evoke any arousal. 2.1.7. Smoothness of interaction Investigating the e!ect of feedback type on the attitudes towards the system, Cassell and Tho H risson (1999) had their subjects communicate with di!erent types of agents about the solar system. In one condition, the agents just provided answers to the user's questions and executed commands. In the second condition, the agents additionally produced emotional responses (such as smiling when addressed by the user or showing confusion when incapable of understanding a user's utterance). In the third condition, the agents provided feedback that underlined the discourse structure (such as turning head and eyes towards the user when expecting a request by the user or averting gaze when executing a command). Although the system's language understanding and generation capabilities were exactly the same in all feedback conditions, subjects rated the smoothness of the interaction as well as the agents' language skills in the discourse-supporting feedback condition signi"cantly higher than in the other two conditions. 2.1.8. Utility Takeuchi and Naito (1995) found that in the context of a card matching game an arrow as a means of visualizing the system's behaviour was numerically perceived as more useful than an animated face. In the same vein, Lester et al. (1997a) report numerically higher utility ratings for a tutoring system with a muted agent than for the same system with a fully expressive agent. However, in the same study, the system's advice was assessed as more useful when the agent was fully expressive as compared to when it was less expressive or muted. Because of qualitative di!erences in the agent's advice, the di!erence in rated utility for the advice is not surprising: whereas the fully expressive agent provided task- and domain-speci"c advice, the muted agent did not supply any advice at all. In contrast to the Lester et al. study, Cassell and Tho H risson (1999) did not vary the amount of information given on a particular topic (the solar system), but only the agent's non-verbal behaviour that accompanied the interaction. In this study, subjects rated the system's utility higher if the agent revealed non-verbal behaviour that underlined the structure of the conversation. Both the agent without non-verbal behaviour as well as the one with emotional responses were experienced as less useful. Nevertheless, even in this study, there was a surplus of information in the discourse-supporting feedback condition. For instance, users were signalled as to whether the system expected an input (by looking at the user) or was busy executing a command (by averting the gaze away from the user).

D. M. DEHN AND S. VAN MULKEN

Only if this surplus of information is in fact speci"c to the animation itself, can it be reasonably interpreted as a positive e!ect of using animated agents in the interface design. Thus, it has to be demonstrated that the surplus of information cannot be easily conveyed by other means (such as meta-text on the monitor). An indirect way of assessing the utility of an animated agent consists in asking the users to rate the di$culty of the task they are faced with in the interaction. If there are di!erences in the assessed di$culty of the task presented by a system with an animated agent and the assessed di$culty of the same task presented by a system without an animated agent, the di!erences can be attributed to presence of the agent. This approach was pursued by van Mulken et al. (1998). The authors found signi"cantly lower task di$culty ratings when technical information was presented by a system with an animated agent as opposed to a system with a pointing arrow. Again, the e!ect of the animated agent appears to be domain-speci"c: if the two kinds of interfaces were used for learning information on "ctitious employees, the animated agent vs. pointing arrow manipulation did not have any e!ect on perceived task di$culty. 2.1.9. Reported attention Koda and Maes asked their subjects (1) whether they paid attention to the animated face in the poker playing game and (2) whether they felt distracted by it. Because the ratings were relatively high for attention to the face and low for distraction by the face, Koda and Maes conclude that &&people were not distracted by the existence of the face'' (p. 192) and that they &&pay attention to the face and engage in the task'' (p. 194). This conclusion does not appear to be valid, though. First, Koda and Maes do not provide any baseline for attention and distraction ratings for a system without a face. Second, it is highly questionable whether subjects have access to their internal processes and thus are able to report meaningfully about them (e.g. Nisbett & Wilson, 1977). In particular, it has been demonstrated that self-reports on the state of attention are completely unrelated to the actual state (Reisberg & McLean, 1985). Hence, the only conclusion to be drawn from the Koda and Maes data is that subjects reported attending to the face and that they did not experience being distracted by it. Similar to Koda and Maes, van Mulken et al. asked their subjects if they felt distracted by an animated agent and if the agent helped them to concentrate on the relevant information. Instead of interpreting the absolute ratings, van Mulken et al. compared the ratings for the agent presentations with the ratings for the arrow presentations. They found no di!erences in perceived distraction, but a signi"cantly higher rating of perceived concentration for the agent condition in the pulley system setting. This again points to the fact that the e!ect of animated agents is domain-speci"c. Table 1 provides a rough summary of the e!ect of animated agents on the user's attitudes.
2.2. THE USER'S BEHAVIOUR WHILE INTERACTING WITH THE SYSTEM

The e!ect of animated agents on the user's style of interaction with the system has so far been investigated with regard to two aspects: the amount of attention directed towards the system or the task and the user's self-presentation to the system. Other measurements that refer more directly to the e$ciency of the interaction (i.e. performance data) rather than to the style of interaction are dealt with in the next section.

THE IMPACT OF ANIMATED INTERFACE AGENTS

TABLE 1 he user1s subjective experience of the system


Variable Intelligence After interaction Before interaction Believiability Lester et al. (1997a) ikeability/Social evaluation Sproull et al. (1996) Koda and Maes (1996) Activity King and Ohya (1996) Sproull et al. (1996) Engagingness/Entertainment Takeuchi and Naito (1995) Koda and Maes (1996) van Mulken et al. (1998) Lester et al. (1997a) Comfortability Sproull et al. (1996) Koda and Maes (1996) ;sability/;tility Takeuchi and Naito (1995) Lester et al. (1997a) van Mulken et al. (1998) Cassell and Tho H risson (1999) Reported Attention Attracts attention Distracts helps concentrate Koda and Maes (1996) Koda and Maes (1996) van Mulken et al. (1998) van Mulken et al. (1998) U(arrow)'U(face) U(muted agent)(U(expressive agent) U(persona)'U(no persona) U(verbal-only), U(verbal/emotional) (U(verbal/structural) Caricature face attracts attention Caricature face does not distract distraction (persona)"distraction(no persona) helps-concentrate(persona)'helpsconcentrate(no persona) C(text with speaking face)(C(text) C(face)'C(no face) E(face)'E(arrow) E(face)'E(no face) E(persona)'E(no persona) E(muted agent)"E(expressive agent) A(geometric shapes)( A(human shapes) A(text with speaking face)(A(text) S(text with speaking face)(S(text) L(face)'L(no face) B(muted agent)(B(expressive agent) Authors Sproull et al. (1996) Koda and Maes (1996) Koda and Maes (1996) King and Ohya (1996) Results I(text with speaking face)"I(text) I(caricature face)"I(no face) I(caricature face)(I(realistic face) I(geometric shapes)(I(human shapes)

Note: Capital letters in the results column are abbreviations of the variable names.

2.2.1. Attention Since attention has the status of a psychological construct, it is not directly observable (e.g. Barkeley, 1996). Thus, the amount of attention the subject directs towards the system or the task can only be inferred indirectly, by using theoretically motivated indicators of attention such as eye contact, reaction times or error rates in the user's responses. Takeuchi and Naito (1995), who investigated eye movements and response times of two opponent (human) players in a virtual card matching game, compared the

10

D. M. DEHN AND S. VAN MULKEN

e!ect of a face in the computer display with that of a 3D arrow. Both displays*the face as well as the arrow*were thought to underline the current state of the game, for instance, by looking at or pointing to the player whose turn it was. Additionally, the face showed a number of emotional expressions, commenting on the current state of the game (e.g. a happy face when the user succeeded in matching two cards). The authors found that the facial display evoked more eye contact than the arrow. This "nding gives evidence that the face attracted more attention than the arrow. Apparently, though, the attraction towards the face distracted the users from their main task, that is, the card matching game: subjects in the face condition needed more time for their reaction (i.e. turning a card) than the subjects in the arrow condition. Takeuchi and Naito conclude from this pattern of results that subjects immediately try to interpret facial displays which in turn &&prevents them from concentrating on the game'' (p. 454). This conclusion*as well as the "nding itself*has to be regarded with some caution, though. First, it remains to be shown that the di!erences in eye contact and response times are not within the limits of random variation. Due to their small number of subjects (N"7), Takeuchi and Naito do not apply any inferential statistics. Second, it is not clear whether the subjects produced longer reaction times because they did not &&concentrate on the game''. Rather, the longer response times might be due to the fact that subjects in the face condition had more cues (i.e. additional facial expressions) to take into account when making their decision. In this way, the greater complexity of the situation might have boosted response times. In the study of Sproull et al. (1996), subjects answered psychological test items presented by an interactive career-counselling system. Here, an interface using textual output was compared with another interface using a talking face in addition to a textual output. In order to measure di!erences in the user's attentiveness to the system, the authors measured the time needed to complete the test items and the number of skipped items. However, in contrast to Takeuchi and Naito, longer response times were assumed to re#ect a higher degree of attentiveness. This should hold because a higher degree of attentiveness should evoke more (time-consuming) thoughts as regards to the &&best'' way of answering the questions. Similarly, a higher degree of attentiveness should also lead subjects to avoid certain personal questions. With regard to the time needed, subjects in the face conditions spent signi"cantly more time to complete the items than did subjects in the text condition. With regard to the number of skipped items, subjects in the face condition left signi"cantly more items unanswered than did subjects in the text condition. Sproull et al. interpret their "nding in support of the hypothesis that people are more attentive to a task presented by a talking-face display than to one presented by a textual display. There are, however, some concerns with regard to the interpretability of the prolonged answer times: the increase in time needed to complete the psychological scales might be at least partially due to a delay between the utterance and the textual display in the face conditions. In the purely textual output condition, there was no such delay. Nevertheless, as Sproull et al. argue, the additional time for auditory presentation cannot fully explain the rise in response time. 2.2.2. Flow of the communication Takeuchi and Nagao (1993) investigated the e!ect of a facial display on the e$ciency of the conversation between a user and a speech dialogue system. In this study, subjects

THE IMPACT OF ANIMATED INTERFACE AGENTS

11

were instructed to question the dialogue system about the functions and prices of computer products. The user's questions as well as the system's responses were given in natural language. In one experimental condition, the system's interface incorporated a 3D image of a young boy showing various facial expressions. In the other condition, the interface did not involve a visual display of a face. Rather, all facial expressions were described verbally on the computer screen. Each subject interacted with both types of displays with the order of interaction being counterbalanced across subjects. In order to measure e$ciency of communication, a &&conversation index'' was formed based on the number of topic shifts and successful answers per time unit. Numerically, the overall di!erence in the conversation indices for the facial vs. non-facial displays was negligible. However, if only the "rst of the two communication sessions was considered, the communication index was bigger for the facial than for the non-facial display. From this "nding, the authors conclude that &&upon "rst contact facial displays clearly help conversation'' (p. 191). This conclusion does not appear compelling though: "rst, the authors did not apply any inferential statistics to their data. For this reason, it cannot be ruled out that the di!erence in the communication scores is due to random #uctuations. Second, unless it is demonstrated that the communication score in fact di!erentiates between e$cient and ine$cient conversations, the choice of this particular index remains rather arbitrary. Cassell and Tho H risson (1999) hypothesized that the e$ciency of the communication with an animated agent*as measured by the number of repetitions and hesitations made by the user as well as communicational overlaps*increases if the agent additionally reveals non-verbal behaviour that indicates the state of the discourse (such as establishing eye contact to signalize turn taking). In contrast, an agent without any non-verbal behaviour as well as one with non-verbal emotional behaviour should yield a less e$cient communication. Corresponding to the expectations, the relative number of repetitions was signi"cantly lower when subjects were provided with non-verbal discourse-structuring feedback. Unexpectedly, though, both the number of hesitations and overlaps were higher in the discourse-supporting condition. Cassell and Tho H risson explain this unexpected "nding by arguing that a bigger number of overlaps might in fact be an indicator for a more &&natural'' interaction. Also, the bigger number of hesitations in the discourse-supporting feedback condition might be due to the fact that users were much more drawn into the communication but had to realize that the current system was not able to keep up with their speed. The validity of these post hoc interpretations remains to be shown, though. 2.2.3. Self-presentation In addition to di!erences in attentiveness to the psychological test items, Sproull et al. also discovered substantial di!erences in the way these items were answered. The test items they presented to their subjects fell into three di!erent psychological scales: the Marlowe}Crowne Social Desirability scale, the Philosophy of Human Nature Altruism scale, and the Texas Social Behaviour Inventory of Self-Worth. Subjects in the face conditions scored higher on both the Social Desirability and the Altruism scale. As opposed to this, there was no di!erence in the scores on the Self-Worth scale. Thus, subjects who interact with a face are more likely to describe themselves in a socially desirable way, and they report themselves as more altruistic than subjects who are just

12

D. M. DEHN AND S. VAN MULKEN

TABLE 2 he user1s behaviour while interacting with the system


Variable Attention Eye contact Response time Skipped items Authors Takeuchi and Naito (1995) Takeuchi and Naito (1995) Sproull et al. (1996) Sproull et al. (1996) Results EC(face)'EC(3D arrow) RT(face)'RT(3D arrow) RT(text with speaking face)' RT(text) SI(text with speaking face)'SI(text) CI(face)"CI(textual descriptions) R(verbal-only), R(verbal/emotional) 'R(verbal/structural) H(verbal-only), H(verbal/emotional) (H(verbal/structural) O(verbal-only), O(verbal/emotional) (O(verbal/structural) SD(text with speaking face)' SD(text) A(text with speaking face)'A(text) SW(text with speaking face)" SW(text)

Flow of communication Conversation index Takeuchi and Nagao (1993) Repetitions Cassell and Tho H risson (1999) Hesitations Overlaps Cassell and Tho H risson (1999) Cassell and Tho H risson (1999)

Self presentation Social desirability scale Sproull et al. (1996) Altruism scale Self-worth scale Sproull et al. (1996) Sproull et al. (1996)

Note: Capital letters in the results column are abbreviations of the variable names.

confronted with a textual output. Table 2 summarizes the empirical "ndings on the user's behaviour during the interaction.
2.3. THE OUTCOME OF THE INTERACTION AS INDICATED BY PERFORMANCE DATA

In the above section, we discussed di!erences in the way a user interacts with a system. The reported measures (such as eye movements and response times) mainly pertained to the construct of attention. In the present section, we are also concerned with behavioural data. These data are thought to re#ect the e!ectiveness of the human}computer interaction with regard to the goal of the interaction process. For instance, for a tutoring system designed to teach, the question to investigate is whether the presence of an animated agent improves learning or not. 2.3.1. Problem solving Lester, Stone, Converse, Kahler and Barlow (1997b) developed a tutoring environment for the domain of botanical anatomy and physiology which incorporated one of "ve clones of an animated agent. These clones di!ered with regard to the type of advice they provided (task-speci"c vs. principle-based) and the modality in which the advice was given (verbal-only or verbal and animated). A muted clone that did not provide any advice at all served as the control condition. In an experimental study, Lester et al.

THE IMPACT OF ANIMATED INTERFACE AGENTS

13

(1997b) had middle school children interact with one of the "ve agent types. During the interaction, the students were faced with a series of design episodes in which a set of environmental speci"cations (e.g. nutrients in the soil and average incidence of sunlight) was given. The students' task was to design a plant that would be likely to survive in the speci"c environment. As a measure of problem-solving ability, Lester et al. collected the number of errors made in the design task. They found a signi"cant interaction of the problem di$culty and agent type on the number of errors: for simple problems, there was no di!erence in error rate between di!erent agents. For complex problems, however, an agent that provided no advice yielded a signi"cantly higher number of errors than all of the advice-giving agents. Lester et al. conclude from their results that an animated agent improves problem solving and that care should be taken that an agent employs various levels of advice and modes of expression. However, this conclusion does not appear to be convincing. First, it is not clear whether the presence of an agent yields any improvements in problem solving beyond those produced by just providing the subject with advice. Second, the results do not suggest that various levels of advice and modes of expression increase problem solving: an equal amount of errors was found for the condition with the fully expressive agent (that provided task-speci"c and principle-based advice both verbally and through animation) and the condition with an agent that provided only verbal task-speci"c advice. Thus, the addition of principle-based advice and animation did not raise the performance over the level obtained with mere verbal task-speci"c advice. In a study by van Mulken et al. (1998) subjects were presented with a series of pulley systems on a computer screen. During the display of a single arrangement, descriptions of the pulley system components were given auditorily. In one group, the auditory descriptions were accompanied by an animated agent that pointed to the components in question. In the other group, the auditory descriptions were complemented with an arrow that pointed to the components in the same way. After the pulley system had vanished, subjects were asked questions on the con"guration and the cybernetics of the arrangement. Following Hegarty and Just (1993), it was assumed that the correct answers to the questions do not merely re#ect memory performance. Rather, they were regarded to indicate problem-solving ability that is dependent on the availability of an adequate mental model. According to the results of van Mulken et al., there was no statistically signi"cant di!erence in the number of correctly answered questions between the subjects who were presented with an animated agent and those who were presented with a graphical symbol. Assuming that the number of correctly answered questions indicates the quality of the mental model that gives rise to the ability to solve problems, the "ndings suggest that the presence of an animated agent did not foster the construction of an adequate mental model and hence did not result in superior problem-solving skills. 2.3.2. Learning In addition to the number of errors in the plant design task, Lester and colleagues also measured students' biological knowledge on the basis of pre- and post-tests. The di!erence in the scores for pre- and post-tests was assumed to re#ect the amount of learning as a result of using the tutoring system. For all experimental conditions, there was a signi"cant di!erence in the pre- and post-test scores. Provided that this di!erence

14

D. M. DEHN AND S. VAN MULKEN

is not just due to some non-experimental factors, the tutoring system succeeded in conveying biological information. However, there was no statistically signi"cant e!ect of the agent type (ranging from muted to fully expressive) on the pre- and post-test di!erences. Numerically, though, the improvements were highest for the fully expressive and lowest for the muted agent. Lester et al. attribute the failure to establish a statistically signi"cant e!ect of agent type on learning performance to the level of di$culty of the test items: the items employed in the test were relatively simple, and*as found for problemsolving e$ciency*di!erences between the agent types might only show up in di$cult tasks. This conjecture was not tested empirically and thus remains speculative. 2.3.3. Memory performance In the second part of their experiment, van Mulken et al. (1998) investigated the e!ect of an animated agent on recall performance for new information. In the study phase of this experiment, subjects were presented with pictures, names, professions and o$ce number of "ctitious employees of a research institute. The employees were again either introduced by the animated agent or were just pointed to by a graphical symbol (the arrow). In the subsequent test phase, subjects were given pictures as cues to which they were asked to produce the name, profession and o$ce number. Again, van Mulken et al. found no di!erence in the recall levels of subjects who were shown the animated agent and those who were shown the graphical symbol. Thus, there was no evidence for a memoryenhancing e!ect of an animated agent in learning new information. Table 3 summarizes the empirical "ndings discussed above.

3. Is there a persona effect?


On the basis of the "ndings reported above, is it appropriate to conclude that there is a positive e!ect of animated agents on the user's attitudes and performance? That is, is there in fact a persona e!ect as some researchers (e.g. Lester et al., 1997a) conclude? To answer this question, we separately discuss the "ndings with regard to the users' attitudes and their behaviour.

TABLE 3 he outcome of the interaction as indicated by performance data


Variable Problem solving van Mulken et al. (1998) Simple problems Lester et al. (1997b) Complex problems Lester et al. (1997b) earning Lester et al. (1997b) Memory van Mulken et al. (1998) PS(persona)"PS(no persona) PS(muted agent)"PS(expressive agent) PS(muted agent)(PS(expressive agent) L(muted agent)"L(expressive agent) M(persona)"M(no persona) Authors Results

Note: Capital letters in the results column are abbreviations of the variable names.

THE IMPACT OF ANIMATED INTERFACE AGENTS

15

3.1. DISCUSSION OF THE FINDINGS ON THE USER'S ATTITUDES

The empirical evidence on the way an animated anthromorphized agent in#uences the user's attitudes towards the system appears to be mixed. The clearest e!ects can be found for the rated entertainment value of a system with an animated agent in comparison to a system without one: generally, a system with an agent is perceived as more entertaining than the one without an agent. There is some evidence, though, that the rise in perceived entertainment does not apply to all systems. If the task or interface is visually attractive from the start, the presence of an animated agent might not make any di!erence. For the social perception (such as perceived intelligence and likeability) of the agent as well as for usefulness and comfortability, the empirical evidence is more inconsistent. In some cases, an anthropomorphized agent is perceived as more believable, comfortable and useful than a non-anthropomorphized agent; in other cases, the opposite holds. For perceived intelligence, there is some evidence suggesting that a human-like visualization is rated as more intelligent than a geometrical one, but there is even more evidence showing that the visualization of the agent does not in#uence the attribution of intelligence. The inconsistent pattern of results might be explained if one takes into account (1) whether the physical appearance is the only information subjects can base their assessment on (2) what particular anthropomorphization is chosen and (3) the domain in which the interaction is set. The information available for making the judgement appears to be especially relevant to all attributes related to social evaluation. It might be true that, in general, some visualizations are assigned a higher level of some attribute such as intelligence than others. However, if the system that uses the visualization demonstrates characteristics that are more diagonistic of the feature to be assessed than its physical appearance, the e!ect of the visualization might disappear. For instance, a human-like "gure might be perceived as more intelligent than a geometrical "gure; however, if a subject played poker against a system, then she would tend to base the intelligence rating for the opponent on his poker-playing skill rather than on the visualization chosen. Social evaluation of the agent is also highly dependent on the particular anthropomorphization used. Likeability ratings do not only depend on whether a face is shown or not, but also on what kind of face is shown. The same holds for the comfortability of subjects interacting with an agent. In this context, the degree of realism is probably an important factor: a high-resolution, fully animated display of a face (accompanied by a computerized voice) probably appears much more like a real social person than a 2D caricature face. Hence, the realistic agent is more likely to be assessed on the basis of standard person perception criteria. Deviations of the animation from &&normal'' human behaviour (e.g. permanent eye contact or a slow unin#ected voice) might then be perceived as &&odd'' resulting in lower ratings for both likeability and comfortability. Further, given a task with high self-involvement, a realistic agent might cause additional social stress to the user (cf. Sproull et al., 1996). The domain in which the interaction is set (i.e. the communicative goal to be achieved in the interaction) is also an important factor in the e!ect of anthropomorphization in human-computer interaction. As mentioned above, the presence of a human-like agent might foster a feeling of uneasiness if the task is highly self-involving (such as answering psychological test items). The e!ect of the domain on the assessed usefulness of an

16

D. M. DEHN AND S. VAN MULKEN

animated agent is even stronger. Whereas, with a technical domain (i.e. pulley systems), subjects perceived the task as less di$cult if there was an animated agent in the presentation rather than a meta-graphical symbol, there was no such di!erence for the learning of social information (i.e. information of "ctitious employees). In a study with a card matching game, an animated agent was assessed as even less useful than a meta-graphical symbol.

3.2. DISCUSSION OF THE FINDINGS ON BEHAVIOURAL DATA

With regard to behavioural data, the "ndings do not support the conclusion that there is a clear persona e!ect. For the user's behaviour during the interaction, there is evidence that a display with a face attracts more attention than a non-facial display. This rise in attention directed towards the face, however, potentially distracts the user's attention away from the task to solve in the interaction (cf. Takeuchi & Naito, 1995). Even in cases in which the facial display fosters task-related internal activity (e.g. checking one's own responses for social desirability), this activity might not be supportive with regard to the goal of interaction (e.g. obtaining an honest and complete description of the user). There is also no clear evidence that the introduction of an animated agent in the interface can enhance the #ow of communication (as measured by the number of topic shifts and successful answers per time unit) compared to interfaces that merely use spoken natural language for input and output (Takeuchi & Nagao, 1993). Similarly, the evidence for positive e!ects of some non-verbal feedback behaviour on the #ow of communication is weak. Although feedback that directly supports the discourse*e.g. establishing eye contact when giving turn*decreases the number of repetitions in a communication, it increases the number of hesitations and conversational overlaps (Cassell & Tho H risson, 1999). With regard to performance data collected after the interaction, it has been argued that animated agents can improve problem-solving e!ectiveness (cf. Lester et al., 1997b, p. 29). However, the empirical results do not support this conclusion. As pointed out above, Lester et al. investigated the e!ect of di!erent levels of advice provided in di!erent modalities, rather than the e!ect of animated agents. Since there was no condition in which some sort of advice was presented by a system without an animated agent, the experiment did not allow for the conclusion that animated agents give rise to better problem-solving skills. A similar criticism applies to the "ndings for learning performance: although Lester et al. provide evidence that their tutoring system succeeded in accomplishing its teaching aims (as indicated by positive pre- and post-test di!erences), they do not show that an animated agent improves learning. The crucial "nding to support this conclusion would have been that the tutoring system with an animated agent gives rise to bigger pre- and post-test di!erences than the one that provides the same amount of advice but does not use an animated agent. This comparison, though, was not included in the study by Lester et al. Cautious conclusions with respect to a persona e!ect for performance data should also be made when taking the "ndings by van Mulken et al. into account: the authors failed to "nd di!erences in both problem solving and learning performance following the interaction with a system that employs an animated agent and a system that uses a meta-graphical symbol. In this study, care was taken to ensure that the same amount of information was given to the user in both conditions, thus providing a fair comparison.

THE IMPACT OF ANIMATED INTERFACE AGENTS

17

3.3. CONCLUSIONS

To date, the literature does not provide evidence for a so-called persona e!ect, that is, a general advantage of an interface with an animated agent over one without an animated agent. One might argue that the e!ect of graphical visualizations on psychological measurements are generally so small in size that statistically signi"cant results are hard to obtain. This, however, is not true. For the use of graphics that accompany texts, a meta-analysis of about 100 experiments has shown a reliable medium-sized e!ect of this manipulation on memory for textual information (cf. Levin, Anglin & Carney, 1987). Nevertheless, it would be oversimplifying to conclude that there is certainly no advantage in using animated agents. In the case of the user's attitudes towards the system, some positive e!ects have been established. Among these is the "nding that a system with an animated agent is largely perceived to be more entertaining than a system without one. For other dimensions, such as utility, likeability and comfortability, the relationship between the kind of interface used (i.e. animated vs. not animated) and the user's ratings appears to be moderated by the kind of animation used and the domain in which the interaction is set. Unfortunately though, the present studies do not enable us to make clear predictions as to what type of animations employed in what type of domain will result in positive attitudes towards the system. For changes in the user's behaviour, the existing evidence is even less compelling. In the light of the small number of studies investigating performance data, though, it would be premature to draw conclusions with regard to the e!ects of animated agents that are too negatively biased. Furthermore, the lack of e!ects that can be uniquely attributed to the use of animated agents is often due to experimental confoundings rather than to the fact that there is no e!ect. A last consideration pertains to the time course in which the experimental studies were set. All studies concerned limited short-term interactions with the systems. It is conceivable, though, that positive e!ects on the user's performance as a result of positive attitudes towards the system take some time to establish. For instance, a greater entertainment value of a tutoring system might lead the student to interact with the system more frequently, thus eventually leading to superior learning achievements. Such an e!ect would not have been detected in any of the studies included in this overview. On the other hand, a longer exposure to an animated agent might wear o! the only reliable e!ect so far, that is, the positive e!ect on entertainment ratings.

4. Suggestions for future research


Systematic and well-conducted investigations on animated interface agents and their e!ect on the user's attitudes and performance are still scarce. There appears to be a need for future studies that (1) overcome the methodological shortcomings of many existing studies and (2) take a more "ne-grained perspective on the e!ect of employing animated agents on the user's motivation and cognition.
4.1. CONFORMING TO METHODOLOGICAL STANDARDS

It seems obvious to demand that further studies should try to conform to the methodological standards of current empirical research. However, the interpretation of many of

18

D. M. DEHN AND S. VAN MULKEN

the existing studies is hampered by the fact that these standards are just not met. We will discuss this claim brie#y for both the experimental manipulation and the measurement of e!ects. With regard to the experimental manipulation, several of the existing studies de"ned the control and experimental conditions in such a way that they di!ered in more than just the dimension under investigation. Consequently, di!erences between the two conditions cannot be attributed exclusively to this dimension. For instance, the condition that served as a control condition in the study by Lester et al. mainly di!ered from the experimental conditions in the amount of advice that was given to the user when interacting with a tutorial system. However, the study was set up to investigate the e!ect of animated pedagogical agents, rather than the e!ect of di!erent amounts of advice (that could be also conveyed without an animated agent). Even in one of the methodologically most sophisticated studies*the study by Sproull et al. (1996)*interpretations are impeded by confoundings. Sproull et al. compared a textual interface with an interface that additionally employed a speaking face. In the latter condition, there was a delay of 1 s between the presentation of the voice and the presentation of the text. In contrast, there was no such delay in the no-face condition. As Sproull et al. admit, di!erences in response times between the two conditions*taken as measures of di!erent levels of attention*are at least partially due to this time lag. Another problem in the realization of the experimental manipulation relates to material e!ects. In some studies, the number of items implementing a particular level of the independent variable is so small that any conclusions concerning the e!ect of the variable are at least questionable. For instance, Koda and Maes investigated the e!ect of the agent's gender on the user's attitudes towards the agent. In order to operationalize gender, the authors used a single instance of a female and a single instance of a male person. By using just one example for each level of the manipulation, the authors are not able to control for random variations in the material that might in#uence the ratings and yet be unrelated to gender (e.g. hair colour or facial expression). With regard to the variables taken in order to measure e!ects, the main problem in the existing studies consists in the failure to demonstrate their validity. For instance, in order to measure the e$ciency of the communication, Takeuchi and Naito calculate a communication index that is based on the number of topic shifts and successful answers per time unit. Without the demonstration that the communication index in fact adequately distinguishes between e$cient and ine$cient communications, though, conclusions that are drawn on its basis are rather worthless. A similar problem can be seen in the ratings presented by Lester et al. (1997a): the authors asked their subjects to assess the system or agent on a 5-point Likert scale. However, the mean ratings they obtained are in some cases rather obscure. For instance, the believability of a muted agent's advice was numerically rated higher than the believability of an agent that gave advice on the task to be solved in the tutorial session. Considering that the muted agent does not provide any advice at all, this "nding should have raised some concerns as regards the interpretability of the ratings. The problem of adequate measures is even more pronounced if the dimension under investigation is a complex psychological construct such as attention and can only be inferred indirectly. The choice of indicators of attention should be guided as much as possible by existing theories (cf. Neumann, 1996). Unfortunately, the gap between observable measures of attention and the hypothetical construct is often not

THE IMPACT OF ANIMATED INTERFACE AGENTS

19

bridged properly leading to inconsistencies in some of the above studies. In their study, Takeuchi and Naito take long response times to be indicators of a low level of attention. Sproull et al., however, interpret long response times as indicators of a high level of attention. There are probably good reasons to propose either relationship, however, the authors fail to present compelling (i.e. theoretically motivated) reasons for their assumptions. In order to avoid the pitfalls described above, future studies should ensure that, in investigating the e!ects of a personi"ed interface, an appropriate experimental design with careful operationalizations of both the experimental manipulation and the e!ect to be measured is employed.

4.2. TAKING A FINE-GRAINED PERSPECTIVE

The research discussed in this paper has shown that a number of factors moderate the way in which an animated interface agent in#uences the user's attitudes towards the system. Among these factors are the type of agent used, the type of information it provides and the domain in which the system is set. For this reason, the simple question as to whether an animated interface agent improves human-computer interaction does not appear to be the appropriate question to ask. Rather, the question to ask is: what kind of animated agent used in what kind of domain in-uences what aspects of the user1s attitudes or performance? For instance, it may be the case that animated agents are only advantageous with regard to the user's attitudes towards the system if they are used in particular domains (such as game playing). The e!ect might be even more restricted in that animated agents are only experienced positively if a particular type of agent is used in combination with a particular domain. For instance, a cartoon-style agent may not be well suited to convey business information, but it may be appropriate to convey information in a game-playing environment. Finally, even further restrictions may apply in that characteristics of the user himself or herself might moderate the e!ect of particular types of agents (see, e.g. Isbister and Nass (1998) on the e!ect of personality match between agent and user, and Lee and Nass (1998) on the e!ect of in-group vs. out-group categorization of the agent). Even if such interactions have been established for some measures (such as preference ratings), it cannot be inferred that they also apply to other measures (such as performance data). The "ndings so far suggest that there are some crucial variables that mediate the e!ect of animated agents on the user's attitudes and cognition. However, the studies that yielded these "ndings were not undertaken in order to systematically investigate moderating variables. Further research is needed in order to "ll this gap. Another question that deserves further investigation concerns the function of the animated agent's behaviour in human-computer interaction. The biggest potential in using animated agents can be seen in the possibility of rendering the interaction multimodal. With an animated agent, certain communicational processes, such as averting gaze, establishing eye contact, and head-, hand- and body gestures can be visualized. With a textual interface, in contrast, such meta-information on the course of the interaction*as well as all other information*has to be formulated explicitly, that is, verbally. In this way, the use of a personi"ed interface might contribute to richer and more natural interaction between human and computer. In the studies reported above, though, most of the agents did not exploit the whole bandwidth of multimodal

20

D. M. DEHN AND S. VAN MULKEN

interaction. In particular, they did not exploit non-verbal behaviour. The only study in which the animated agent showed non-verbal (i.e. emotional and discourse-supporting) behaviour that resembled the behaviour of a human interaction partner is the study by Cassell and Tho H risson (1999). In our opinion, an animated agent can only be expected to improve human-computer interaction, if it shows some behaviour that is functional with regard to the system's aim. For instance, if the goal is to render the interaction smoother, then the personi"ed agent has to reveal behaviour (such as non-verbal behaviour like averting gaze, smiling, etc.) that can be assumed to support the process of interaction without being a potential source of distraction (cf. Cassell & Thorisson, 1999). In contrast, if the goal is to render a tutorial system more entertaining, then it might be even sensible to employ an animated agent that indulges in #ying, bungee-jumping, acrobatics, and so on (cf. Lester et al., 1997a). As mentioned above, the lack of an agent's e!ect on the user's performance might be due to the fact that the existing studies only investigated (relatively) short-term e!ects that emerged after some clearly de"ned interaction sessions. Given the fact that an animated agent raises the entertainment value of the system (and given that this e!ect does not vanish with longer exposure to the system), the user should be more motivated to actively seek contact with the system. More and longer interactions with the system should eventually yield better performance in terms of the system's goal. Thus, future studies should focus on potential long-term e!ects of animated agents. Finally, it can also be asked under what circumstances an animated agent does not contribute to a smooth and e!ortless interaction but rather distracts the user from their current task. For instance, in some applications, the agent's behaviour neither maps onto system behaviour nor conveys any other information that is relevant to the user. This type of behaviour is often used to render the agent more lively and is exhibited during idle time (consider, for instance, Microsoft's paper clip assistant). Potentially, though, such behaviour can also impede the interaction. For instance, the movements might make the user attend to the agent although at this time it does not provide any information. Moreover, the user might wonder what he or she did that made the agent blink its eyes or behave in some other way. In any case, the idle-time behaviour would require cognitive resources without conveying any information. On the other hand, idle-time behaviours may indeed render an agent more lively, and thus more human-like. This might have the desirable e!ect that particularly novice computer users lose their fear of computers and be more inclined to interact with the computer spontaneously*as they would do with fellow human beings.
This research was partly supported by the European Community under ERB 4061 PL 97-0808. We thank Elisabeth Andre H and three anonymous reviewers for valuable comments on a previous version of this paper.

References
ANDRE D , E. & RIST, T., Eds. (1997). Proceedings of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent. Nagoya, Japan, 25 August.

THE IMPACT OF ANIMATED INTERFACE AGENTS

21

ANDRE D , E. (1999). Animated interface agents (special issue). Applied Arti,cial Intelligence, 13. BALL, G., LING, D., KURLANDER, D., MILLER, J., PLUGH, D., SKELLY, T., STANKOSKY, A., THIEL, D., DANTZICH, M. V. & WAX, T. (1997). Lifelike computer characters: the persona project at microsoft research. In J. M. BRADSHAW, Ed. Software Agents, pp. 191}222. Menlo Park, CA: AAAI/MIT Press. BARKELEY, R. A. (1996). Critical issues in research on attention. In G. R. LYON & N. A. KRASUEGOV, Eds. Attention, Memory, and Executive Function, pp. 45}56. Baltimore: Paul Brookes. BESKOW, J. & MCGLASHAN, S. (1997). Olga*a converstational agent with gestures. Proceedings of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent. Nagoya, Japan, 25 August. CASSELL, J., PELACHAUD, C., BADLER, N., STEEDMAN, M., ACHORN, B., BECKETT, T., DOUVILLE, B., PREVOST, S. & STONE, M. (1994). Animated conversation: rule-based generation of facial expression, gesture, and spoken intonation for multiple conversational agents. Proceedings of the SIGGRAPH Conference on Computer Graphics, pp. 413}420. CASSELL, J. & THO D RISSON, K. R. (1999). The power of a nod and a glance: envelope vs. emotional feedback in animated conversational agents. Applied Arti,cial Intelligence, 13, 519 } 538. ELLIOT, C., RICKEL, J. & LESTER, J. C. (1997). Integrating a!ective computing into animated tutoring agents. Proceedings of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent. pp. 113}121. Nagoya, Japan, 25 August. HEGARTY, M. & JUST, A. (1993). Constructing mental models of machines from text and diagrams. Journal of Memory and anguage, 32, 717}742. HUARD, R. E. & HAYES-ROTH, B. (1997). Character mastery with improvisational pupets. Proceedings of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent, pp. 85}89. Nagoya, Japan, 25 August. ISBISTER, K. & NASS, C. (1998). Personality in conversational characters: building better digital interaction partners using knowledge about human personality preferences and perceptions. Proceedings of the 1st Workshop on Embodied Conversational Characters (WECC+98), pp. 103}111. Lake Tahoe, CA, 12}15 October. KING, W. J. & OHYA, J. (1996). The representation of agents: anthropomorphism, agency, and intelligence. In R. BILGER, S. GUEST & M. J. TAUBER, Eds. Human Factors in Computing Systems: CHI+96 Electronic Conference Proceedings. ACM. Available via http://info.acm.org/ sigchi/chi96/proceedings/shortpap.htm. KODA, T. & MAES, P. (1996). Agents with faces: the e!ect of personi"cation. Proceedings of the 5th IEEE International Workshop on Robot and Human Communication (RO-MAN+96), pp. 189}194. LAUREL, B. (1990). Interface agents: metaphors with character. In B. LAUREL, Ed. he Art of HCI Design, pp. 355}365. Reading, MA: Addison-Wesley. LEE, E.-J. & NASS, C. (1998). Does the ethnicity of a computer agent matter? An experimental comparison of human-computer interaction and computer-mediated communication. Proceedings of the 1st Workshop of Embodied Conversational Characters (WECC+98), pp. 123}128. Lake Tahoe, CA, 12}15 October. LESTER, J. C., CONVERSE, S. A., KAHLER, S. E., BARLOW, S. T., STONE, B. A. & BHOGAL, R. S. (1997a). The persona e!ect: a!ective impact of animated pedagogical agents. In S. PEMBERTON, Ed. Human Factors in Computing Systems: CHI+97 Conference Proceedings, pp. 359}366. New York: ACM Press. LESTER, J. C. & STONE, B. A. (1997). Increasing believability in animated pedagogical agents. Proceedings of the 1st International Conference on Autonomous Agents, pp. 16}21. Marina del Rey, CA. LESTER, J. C., STONE, B. A., CONVERSE, S. A., KAHLER, S. E. & BARLOW, S. T. (1997b). Animated pedagogical agents and problem-solving e!ectiveness: a large-scale empirical investigation . In B. DU BOULAY & MIZOGUCHI, Eds. Proceedings of the 8th World Conference on Artixcial Intelligence in Education, pp. 23}30. Kobe, Japan: IOS Press. LEVIN, J. R., ANGLIN, G. J. & CARNEY, R. N. (1987). On empirically validating functions of pictures in prose. In D. M. WILLOWS & H. A. HOUGHTON, Eds. he Psychology of Illustration, Vol. 1, pp. 51}85. New York: Springer.

22

D. M. DEHN AND S. VAN MULKEN

MAES, P. (1994). Agents that reduce work and information overload. Communications of the ACM, 37, 31}40. NASS, C., MOON, Y., FOGG, B., REEVES, B. & DRYER, D. C. (1995). Can computer personalities be human personalities? International Journal of Human-Computer Studies, 43, 223}239. NASS, C., STEUER, J. & TAUBER, E. R. (1994). Computers are social actors. In B. ADELSON, S. DUMAIS & J. OLSON, Eds. Human Factors in Computing Systems: CHI+94 Conference Proceedings, pp. 72}77. New York: ACM Press. NEUMANN, O. (1996). Theories of attention. In O. NEUMANN & A. F. SANDERS, Eds. Handbook of Perception and Action, <ol. 3: Attention, pp. 389}446. London: Academic Press. NISBETT, R. & WILSON, T. (1977). Telling more than we know: verbal reports on mental processes. Psychological Review, 84, 231}259. NORMAN, D. A. (1994). How might people interact with agents? Communications of the ACM, 37, 68}71. REISBERG, D. & MCLEAN, J. (1985). Meta-attention: do we know when we are being distracted? he Journal of General Psychology, 112, 291}306. RICKEL, J. & JOHNSON, W. L. (1999). Animated agents for procedural training in virtual reality: perception, cognition, and motor control. Applied Arti,cial Intelligence, 13, 343 } 382. RIST, T., ANDRE D , E. & MU G LLER, J. (1997). Adding animated presentation agents to the interface. In J. MOORE, E. EDMONDS & A. PUERTA, Eds. Proceedings of the International Conference on Intelligent User Interfaces, pp. 79}86. Orlando, FL, New York: ACM Press. SHNEIDERMAN, B. & MAES, P. (1997). Direct manipulations vs. interface agents: excerpts from debates at IUI'97 and CHI'97. Interactions, 4, 42}61. SPROULL, L., SUBRAMANI, M., KIESLER, S., WALKER, J. H. & WATERS, K. (1996). When the interface is a face. Human}Computer Interaction, 11, 97}124. SULLIVAN, J. & CASSELL, J. (1998). Proceedings of the 1st Workshop on Embodied Conversational Characters (WECC+98). Lake Tahoe, CA, 12}15 October. TAKEUCHI, A. & NAGAO, K. (1993). Communicative facial displays as a new conversational modality. Proceedings of the Conference on Human Factors in Computing Systems, INTERCHI+93, pp. 187}193. Amsterdam, The Netherlands. TAKEUCHI, A. & NAITO, T. (1995). Situated facial displays: towards social interaction. In I. KATZ, R. MACK, L. MARKS, M. B. ROSSON & J. NIELSEN, Eds. Human Factors in Computing Systems: CHI+95 Conference Proceedings, pp. 450}455. New York: ACM Press. TRAPPL, R. PETTA, P. Eds. (1997). Creating Personalities for Synthetic Actors: Toward Autonomous Personality Agents. Berlin: Springer. VAN MULKEN, S., ANDRE D , E. & MU G LLER, J. (1998). The persona e!ect: how substantial is it? In H. JOHNSON, L. NIGAY & C. ROAST, Eds. People and Computers XIII: Proceedings of HCI+98, pp. 53}66. Berlin: Springer. WALKER, J. H., SPROULL, L. & SUBRAMANI, R. (1994). Using a human face in an interface. In B. ADELSON, S. DUMAIS & J. OLSON, Eds. Human Factors in Computing Systems: CHI+94 Conference Proceedings, pp. 85}91. New York: ACM Press. WILSON, M. (1997). Metaphor to personality: the role of animation in intelligent interface agents. Proceedings of the IJCAI-97 Workshop on Animated Interface Agents: Making them Intelligent. Nagoya, Japan, 25 August. WRIGHT, P., MILROY, R. & LICKORISH, A. (1999). Static and animated graphics in learning from interactive texts. European Journal of Psychology of Education. (Special Issue on Visual Learning with New Technologies) (in press). ZAJONC, R. (1965). Social facilitation. Science, 149, 269}274.

Anda mungkin juga menyukai