Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 7-Day Trial for You or Your Team.

Learn More →

Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the Uncanny Valley

Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the... Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the Uncanny Valley Markus Thaler Stephan Schlogl ¨ Aleksander Groth Management, Communication & IT Management, Communication & IT Management, Communication & IT MCI – The Entrepreneurial School MCI – The Entrepreneurial School MCI – The Entrepreneurial School Innsbruck, Austria Innsbruck, Austria Innsbruck, Austria mark.thaler@mci4me.at stephan.schloegl@mci.edu aleksander.groth@mci.edu Abstract—Visual appearance is an important aspect influenc- where a too realistic appearance is found to trigger feelings ing the perception and consequent acceptance of Embodied Con- of eeriness in the human interlocutor – a phenomenon that versational Agents (ECA). To this end, the Uncanny Valley theory th was already described in the early 20 century by Jentsch, contradicts the common assumption that increased humanization observing interactions with automatic dolls [10]. The heron of characters leads to better acceptance. Rather, it shows that built Uncanny Valley effect demonstrates that the affinity anthropomorphic behavior may trigger feelings of eeriness and rejection in people. The work presented in this paper explores with a given object is based on a non-linear relationship whether four different autonomous ECAs, specifically build for with the human-likeness of the respective object [11]. Our a European research project, are affected by this effect, and how study examines to which extent this effect can be found they compare to two slightly more realistically looking human- with the Embodied Conversational Agents (ECA) used in the controlled, i.e. face-tracked, ECAs with respect to perceived European research project EMPATHIC . In particular, our humanness, eeriness, and attractiveness. Short videos of the ECAs in combination with a validated questionnaire were used to goal was to investigate the following research question: investigate potential differences. Results support existing theories highlighting that increased perceived humanness correlates with How do human-driven, i.e. face-tracked, ECAs (so-called increased perceived eeriness. Furthermore, it was found, that avatars) compare to autonomous ECAs with respect to their neither the gender of survey participants, their age, nor the perceived humanness, eeriness and attractiveness? sex of the ECA influences this effect, and that female ECAs are perceived to be significantly more attractive than their male II. RELATED WORK counterparts. Index Terms—Embodied Conversational Agents, Avatars, Un- Although recent years have seen great progress in the devel- canny Valley, Humanness, Eeriness, Attractiveness opment of speech and language-based user interfaces, people still feel uneasy when ‘talking’ to an artificial entity [12], [13]. I. INTRODUCTION One reason for this may be rooted in the fact, that a large For decades, research and industry have been struggling to part of human communication lies outside the scope of pure build human-like computing systems, i.e. digital assistants, language. ECAs aim to ameliorate this shortcoming through an that can be operated via natural interaction modalities (e.g. additional emotional interaction channel. That is, they not only natural language, gestures, mimics, etc.) and are capable support speech-based interaction [13], [14], but also feature of understanding and expressing emotions. Due to past some sort of visual and/or physical appearance [15], which al- developments in artificial intelligence and natural language lows for the generation of facial expressions. Consequently, the processing, and the increasing uptake of digital assistants human interlocutors’ abilities to predict an agent’s behaviour in the consumer sector, this gap between science fiction improve [16]. In addition, the embodiment lets an agent appear and reality is, however, gradually closing. With respect more aesthetic and life-like [4], [17] as well as potentially to application areas for these types of intelligent systems, more intelligent [8]. Yet, improvement stops once the agent research is often looking at the social and health care crosses the so-called Uncanny Valley (cf. Fig. 1). context, where people may benefit from support in various A. The Uncanny Valley quotidian tasks [1]–[3]. Here, the question as to which social characteristics a digital agent should or should not exhibit in According to Mori [11], the Uncanny Valley (UV) shows these settings, is continuously being researched [4]–[6]. One the affinity a human observer has towards an object and how of these issues concerns the right embodiment of respective such changes when the human likeness of the object increases. systems, and how such may influence their credibility [7]–[9]. Humanization, in particular, may pose unwanted side-effects http://www.empathic-project.eu/ arXiv:2104.11043v1 [cs.HC] 22 Apr 2021 B. Anthropomorphism The anthropomorphism phenomenon is one crucial aspect related to the UV theory, stating that human characteristics, motivations, intentions and emotions may be attributed to non- human entities, even if these have only slightly human-like traits [25]. Duffy [7, p. 180] describes this as “attributing cog- nitive or emotional states to something based on observation in order to rationalise an entity’s behaviour in a given social environment”. In other words, anthropomorphism is consid- ered a psychological process helping humans rationalize ob- servations. Guthrie [26] describes the respective procedure in Fig. 1. The Uncanny Valley according to Mori [11] four consecutive steps: (1) one is confronted with an unknown object which initially triggers shallow conclusions; (2) this initial conclusions are then combined with and expanded upon past experiences; (3) the most probable template describing Ho & MacDorman [18] describe this as a graph illustrating a the object is selected; and (4) since the ‘human’ template non-linear relationship between the human resemblance of an is most commonly chosen, the observed characteristics are anthropomorphic entity and a viewer’s emotional response. As compared to those of humans. As this process is based on can be seen in Fig. 1, Mori chose industrial robots as a starting non-rational thinking, humanization can be perceived nega- point, since they exhibit little resemblance to humans. At the tively [27]. Contrary to this, in Human-Computer Interaction opposite side of the valley one then finds for example realistic research and design, anthropomorphism is often considered hand prostheses, as they are rather difficult to distinguish from a design guideline that helps build appealing user interfaces real human hands. According to Seyama & Nagayama [19], (UI), and aims at triggering rather pleasant experiences in the UV effect can be shown with both physical and virtual users [4]. While humans often anthropomorphise, even in objects. In an experiment with computer-generated characters cases where an interface exhibits hardly any human traits, MacDorman and colleagues, for example, found that faces it has been shown that the context and given task setting with photo-realistic textures and a high degree of detail tend to influence the perception and consequent acceptance of this trigger eeriness. Characters with a lower photo-realism level human-likeness [28]. In social tasks, for example, a human- (i.e., 75%), however, were deemed the least uncanny by human like agent seems acceptable, whereas in other settings an viewers [20]. Hence, it may be argued that the UV theory anthropomorphic appearance may be less expected, and con- contradicts the view that a robot or embodied agent should sequently trigger feelings of uneasiness [29]. resemble a human being as close as possible [4]. The second part of the UV theory deals with motion. C. ECAs in Social and Health Care Settings According to Mori [11], the amplitude of the curve increases ECAs may help simplify human-technology interaction. when the object is in motion. In an experiment with avatars For seniors, in particular, accessibility to technology can be (i.e. human-controlled agents), it was found that motion influ- improved [3], [30]. One area that sees an ongoing uptake of ences familiarity and perceived attractiveness of the agent [21]. these types of intelligent systems is the social and health care Although, it has to be underlined that both familiarity and sector, where ECAs are increasingly used to support everyday perceived attractiveness are subject to personal taste. For tasks of people [31]. However, as already outlined above, the example, it was shown that even very human-like characters visual appearance of these ECAs poses certain challenges with may be deemed repelling [19]. Similarly, Jentsch notes in respect to their human-likeness and the consequent effects on his definition of uncannyness, that a realistic technological people’s feelings of eeriness. Thus, the goal of our work was to stimulus can trigger a feeling of eeriness in people [10]. There explore how ECAs specifically created for health related tasks, are several theories where this sensation of eeriness might i.e. the ECAs built for the EMPATHIC project, are perceived come from [19]. Bartneck et al. attribute it to the framing by the general population. theory [22], which states that when observing new things, III. M ETHODOLOGY humans refer to past experiences, so-called frames. This gives rise to expectations that may not always be met. With human- The goal was to evaluate six different ECAs (cf. Figures 2 like robots, for example, the ‘human frame’ is selected, but & 3). Four autonomous agents, which were created for the when seeing the robot’s mechanical movements, which may EMPATHIC project, i.e. Lena, Alice, Christian and Adam, resemble those of a sick or injured person, it is often the case and two human-driven/face-tracked agents (so-called avatars), that a feeling of discomfort arises. Another trigger for this i.e. Sophie and Michael, which were built separately, so as eeriness may be found in the skin discoloration of artificial to serve as a control group (note: for these two agents we characters, which reminds people of death and thus may evoke 2 used the Reallusion Software Suite , which led to slightly respective fears [14], [23], [24]. Here one might also draw a connection to the psychological process of anthropomorphism. https://www.reallusion.com/iclone/ perceived attractiveness dependent on the sex of the ECA (cf. Esposito et al. [33]). H5a: There is a significant positive correlation between a participant’s age and the perceived eeriness of ECAs (cf. Yaghoubzadeh et al. [34]). H5b: There is a significant negative correlation between a participant’s age and the perceived attractiveness of ECAs (cf. Yaghoubzadeh et al. [34]). H6: Human-driven ECAs (i.e., avatars) exhibit a higher level of perceived eeriness than autonomous ECAs. In order to evaluate these hypotheses and consequently the extent to which the tested ECAs may fall into the UV, we exposed participants to an eight-second video of an agent and subsequently asked them to rate the agent’s perceived level of humanness, eeriness and attractiveness according to the following 21 UV semantic differential effect scales proposed Fig. 2. Autonomous ECAs built for the EMPATHIC project by MacDorman & Ho [35] (note: in brackets we provide the German translation of the terms the way they were used in the questionnaire): Humanness: 7-point semantic differentials on – inanimate $ living (ge: unbelebt/lebendig) – synthetic $ real (ge: synthetisch/echt) – mechanical movement $ biological movement (ge: mech. Bewegungen/nat. Bewegungen) – human-made $ humanlike (ge: von Menschen gemacht/menschenahnlich) – without definite lifespan $ mortal Fig. 3. Human-driven ECAs (built using the Reallusion Software Suite) (ge: ohne definitive Lebensdauer/sterblich) – artificial $ natural (ge: kunstlich/nat ¨ urlich) ¨ Eeriness: 7-point semantic differentials on more realistic faces and facial expressions than what was – dull $ freaky (ge: eintonig/ausgef ¨ allen) achieved by the project ECAs). All six ECAs exhibited similar – predictable $ eerie (ge: abschatzbar/unheimlich) ¨ characteristics with respect to Ring et al.’s classification of – plain $ weird (ge: schlicht/sonderbar) anthropomorphic UIs [29]. That is, all were human(-like), – ordinary $ supernatural mimicking the same ethnicity and similar age and wearing (ge: gewohnlich/außernat ¨ urlich) ¨ similar clothing. Only the level of realism distinguished the – boring $ shocking (ge: langweilig/schockierend) project ECAs (i.e., Lena, Alice, Christian and Adam) from – uninspiring $ spine-tingling the slightly more realistic Reallusion ECAs (i.e., Sophie and Michael). In accordance with the above discussed literature (ge: uninspirierend/elektrisierend) we thus formed the following hypotheses to be tested: – predictable$ thrilling (ge: vorhersehbar/mitreißend) – bland $ uncanny (ge: nichtssagend/untypisch) H1: Human-driven ECAs (i.e., avatars) exhibit a higher – unemotional $ hair-raising level of perceived humanness than autonomous ECAs (cf. (ge: emotionslos/furchterregend) MacDorman et al. [20]). – familiar $ uncanny (ge: vertraut/unheimlich) H2: The perceived level of humanness has a significant Attractiveness: 7-point semantic differentials on influence on the perceived level of eeriness (cf. McDon- – ugly $ beautiful (ge: hasslich/sch ¨ on) ¨ nell et al. [21]). – repulsive $ agreeable (ge: abstoßend/ansprechend) H3a: There is a significant difference between genders – crude $ stylish (ge: geschmackslos/stylisch) when it comes to the perceived level of eeriness (cf. – messy $ sleek (ge: unordentlich/gepflegt) Tinwell & Sloan [32]). – unattractive $ attractive (ge: unattraktiv/attraktiv) H3b: There is a significant difference between genders when it comes to the perceived level of attractiveness This procedure was repeated for each of the six ECAs. (cf. Tinwell & Sloan [32]). Finally, participants were asked to provide some additional H4a: There is a significant difference with respect to demographic information. We particularly targeted German- perceived eeriness dependent on the sex of the ECA (cf. speaking participants so as to prevent any potential cultural Esposito et al. [33]). bias. To this end, all scales and questions were translated. A H4b: There is a significant difference with respect to German-speaking lecturer of English as well as a bilingual TABLE I R ELIABILITY OF S CALES Stimulus Humanness ( ) Attractiveness ( ) Eeriness ( ) Lena 0.871 0.896 0.827 Adam 0.925 0.896 0.856 Sophie 0.937 0.932 0.864 Christian 0.956 0.910 0.878 Alice 0.948 0.921 0.878 Michael 0.947 0.951 0.894 Overall 0.957 0.926 0.950 Fig. 4. Mean Values of Perceived Humanness for Each of the Shown ECAs colleague were consulted to improve translation quality. In TABLE II addition, all scales were tested for reliability before further ANOVA - POST-HOC S CHEFFE evaluations began. Participation in the study was voluntary Stimuli Humanness (p) Eeriness (p) (note: three EUR 10,- Amazon Gift Vouchers were raffled Alice ! Adam 0.0015** 0.8295 off among the participants). Furthermore, questionnaire and Christian ! Adam 0.9994 0.9988 respective procedure were evaluated by MCI’s research ethics Lena ! Adam 0.9995 0.9976 Michael ! Adam 0.0000*** 0.0000*** group for accordance with ethical guidelines concerning re- Sophie ! Adam 0.0000*** 0.0903 search with human participation. Christian ! Alice 0.0066** 0.5834 Lena ! Alice 0.0064** 0.5441 IV. R ESULTS Michael ! Alice 0.3798 0.0068** Sophie ! Alice 0.5038 0.7548 A total of n = 215 participants (150 females) completed Lena ! Christian 1.0000 1.0000 the above described procedure, approx. 50% (i.e. 108) of Michael ! Christian 0.0000*** 0.0000*** whom were students (predominantly business and information Sophie ! Christian 0.0000*** 0.0266* Michael ! Lena 0.0000*** 0.0000*** systems but also other fields such as pharmacy and physics). Sophie ! Lena 0.0000*** 0.0219* The overall age distribution (74%  30 years; 21% 31 Sophie ! Michael 1.0000 0.3367 60 years; 5% > 60 years) shows a strong right-bound skewness of 1.766 (Mean = 31:04; Median = 25). Scale reliability for humanness, eeriness and attractiveness were first in perceived humanness (p = 0:014) attributed to the type of tested individually for each stimulus. Then each dimension control, i.e. autonomous vs. human-controlled. An ANOVA was analyzed in its entirety using Cronbach’s . Table I shows including the Scheffe ´ post-hoc test generally confirmed this that the overall scale reliability was excellent ( = 0:95) significantly higher level of perceived humanness with the and when focusing on internal stimuli only, it was still good Reallusion ECAs, yet further showed that also Alice (the most ( > 0:70) [35]. Given this reliability of scales we were human-like project ECA) scored significantly higher than the able to proceed, evaluating and consequently comparing the other project ECAs (cf. Table II). All in all, however, we may six ECAs with respect to their perceived humanness, eeriness, argue that our data supports H1. and attractiveness. Respective results are summarized in the following sub-sections. B. Eeriness A. Humanness In order to explore the UV effect, ECAs were ordered Fig. 4 shows the mean values of the six ECAs with respect according to their human similarity scores and then evalu- to humanness (note: we calculated the mean over all the ated with respect to their perceived level of eeriness. Fig. 5 scales concerning humanness described in Section III where shows that after initial improvements eeriness starts increasing values close to 1 signify low humanness and values close to with increasing humanness. That is, the project ECAs Adam 7 signify high humanness). To this end, the project ECAs (x  = 3:19; SD = 0:94), Lena (x  = 3:14; SD = 0:80) Adam (x  = 2:58; SD = 1:22), Lena (x  = 2:63; SD = 1:09) and Christian (x  = 3:15; SD = 0:95) had lower eeriness and Christian (x  = 2:63; SD = 1:36) were perceived quite scores than Alice (x  = 3:32; SD = 0:89). Similarly, the similarly. Alice, however, was perceived as more human-like Reallusion ECAs Sophie (x  = 3:46; SD = 0:90) and Michael (x  = 3:15; SD = 1:39). The two Reallusion ECAs Sophie (x  = 3:67; SD = 1:01) scored rather negatively, with Michael (x  = 3:41; SD = 1:38) and Michael (x  = 3:44; SD = 1:48) moving even further away from Sophie and again showing were also perceived similarly and more human-like than any the greatest variability in people’s perception of eeriness. A of the project ECAs. Yet, Michael’s ratings showed the highest T-test for independent samples showed that this perception variability. Although, human similarity exhibited generally a was independent of people’s gender (F-test for gender variance higher standard deviation than the other dimensions. equality: p = 0:752), for which H3a had to be rejected (p = A T-test for combined samples between the most human- 0:655). Similarly with respect to the agents’ sex no significant like project ECA (i.e., Alice) and the least human-like Re- difference in perceived eeriness was found (p = 0:411). Hence, allusion ECA (i.e., Sophie) points to a significant difference also H4a was rejected. Fig. 5. Mean Values of Perceived Eeriness for Each of the Shown ECAs Fig. 6. Mean Values of Perceived Attractiveness for Each of the Shown ECAs Looking at a connection between humanness and eeriness, a Finally, investigating a potential negative relation between Pearson correlation analysis points to a medium to strong pos- the participants’ age and the perceived attractiveness of ECAs, itive correlation between those two variables (r = 0:573; p = no significant correlation was found (r = 0:095; p = 0:164), 0:000). The subsequently performed linear regression shows for which also H5b had to be rejected. that perceived humanness is able to explain 32.9% of the vari- ance in perceived eeriness (R = 0:329; Beta = 0:573; p = V. S UMMARY, L IMITATIONS AND OUTLOOK 0:000). Hence, H2 is supported by the data. The goal of our study was to examine six different ECAs With respect to H5a, we explored a potential positive concerning characteristics of the UV. According to the clas- relation between a participant’s age and the ECA’s perceived sification by Ring and colleagues [29] we expected a more eeriness. The performed Pearson correlation analysis, however, realistic rendering style to perform better than a less realistic shows a weak but significant negative connection between rendering style. Consequently, we evaluated the ECAs with re- those two variables (r = 0:150; p = 0:014). Consequently, spect to their perceived humanness, attractiveness and eeriness. H5a had to be rejected. The results of our evaluation do not fully confirm our expecta- Finally, a T-test for combined samples pointed to a sig- tions, yet they are in line with previous work. That is, we found nificant difference between the most eerie project ECA (i.e., a positive correlation between an ECA’s human similarity and Alice) and the least eerie Reallusion ECA (i.e., Sophie), its perceived eeriness. Analogous to the UV theory, this means supporting the assumption outlined by H6 (p = 0:024). The that with increasing realism, the viewer’s affinity with an ECA subsequently performed ANOVA and Scheffe ´ post-hoc test decreases. Age-related connections could only be confirmed to confirmed this significantly higher level of perceived eeriness a limited extent, showing a slight negative correlation with the for Michael (i.e., significantly higher eeriness scores than all participant’s eerieness ratings. The participant’s gender had project ECAs) and partly for Sophie (i.e., significantly higher no effects on the ratings, the sex of the stimuli, however, eeriness scores than Christian and Lena). Overall, we thus influenced their perceived attractiveness, with female agents may argue that our data supports H6 in that the perceived being perceived significantly more attractive than male ones. eeriness of the two Reallusion ECAs (i.e., the avatars) was Although these results generally confirm previous work, we rated significantly higher than the perceived eeriness of the would like to point to a number of limitations of our study. four project ECAs (cf. Table II). First, while our human-driven ECAs were indeed perceived C. Attractiveness more human-like than the autonomous ECAs built for the EMPATHIC project, their level of humanness was still rather Attractiveness was generally rated higher than the other low (i.e., x  = 3:41 and x  = 3:44 respectively). Such may two dimensions. Lena (x  = 4:61; SD = 1:01) and Sophie have had an affect on the significance of results, in particular (x  = 4:40; SD = 1:21) were rated the most attractive, with respect to the origin of perceived eeriness. Second, the Adam the least attractive (x  = 3:43; SD = 1:07), and Alice (x  = 4:16; SD = 1:17), Michael (x  = 4:07; SD = 1:35) and used questionnaire was rather long and monotonous, repeating Christian (x  = 3:96; SD = 1:08) lay somewhere in between. the same sections (i.e. MacDorman & Ho’s 21 semantic Also here the participants agreed the least with the evaluation differential effect scales) six times, which may have influenced of Michael. In Fig. 6 the ECAs were ranked according to their people’s responses. Third, the negative correlation we found perceived humanness. Looking at the graph, it is noticeable with respect to participants’ age and their provided eeriness that the female agents were consistently classified as more ratings was very weak (cf. H5a). Finally, a more homogeneous attractive. A T-test on the mean values of the ECA’s sex sample may have potentially provided better insights. That is, supports this assumption, showing a significant difference in although in general our data did not point to any irregularities, terms of perceived attractiveness (p = 0:000). Consequently, the demographic distribution was not ideal. The majority of H4b is supported by the data. From a participants’ point respondents were students younger than 30 years, and more of view, however, no gender differences with respect to the than two-thirds were female. perceived attractiveness of ECAs was found (p = 0:278). In conclusion, this study has confirmed previous findings Hence H3b had to be rejected. with respect to the UV theory; i.e. the more human-like an ECA appears the greater its produced feeling of eeriness. An [14] L. Ciechanowski, A. Przegalinska, M. Magnuski, and P. A. Gloor, “In the shades of the uncanny valley: An experimental study of human-chatbot additional, not negligible result of our study, is the translation interaction,” Future Generation Comp. Syst., vol. 92, pp. 539–548, 2018. and validation of MacDorman & Ho’s UV scales into German, [15] S. Franklin and A. Graesser, “Is it an agent, or just a program?: A tax- which may allow for additional future studies in German- onomy for autonomous agents,” in Intelligent Agents III Agent Theories, Architectures, and Languages, J. P. Muller ¨ , M. J. Wooldridge, and N. R. speaking countries. Those studies may, for example, focus Jennings, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 1997, on the level of realism that is accepted with ECAs. To this pp. 21–35. end an interesting question to investigate may be whether [16] T. Koda and P. Maes, “Agents with faces: the effect of personification,” IEEE Int. Symp. on Robot and Human interactive Communication, pp. ECAs could be made so realistic and credible that they would 189 – 194, 1996. ‘climb up’ the other end of the UV. Also, it may be worth [17] J. Cassell, “Embodied conversational agents: Representation and intel- exploring whether female ECAs are consistently rated more ligence in user interfaces.” AI Magazine, vol. 22, pp. 67–84, 2001. [18] C.-C. Ho and K. MacDorman, “Revisiting the uncanny valley hypothe- attractive than male ones. Finally, concerns regarding privacy sis: Developing and validating an alternative to the godspeed indices,” and ethics connected to the use of ECAs would require Computers in Human Behavior, vol. 26, pp. 1508–1518, 2010. additional research. In particular, if their potential application [19] J. Seyama and R. Nagayama, “The uncanny valley: Effect of realism on the impression of artificial human faces,” Presence, vol. 16, pp. 337–351, domain is focusing on social and health care settings. [20] K. MacDorman, R. Green, C.-C. Ho, and C. Koch, “Too real for ACKNOWLEDGMENT comfort? uncanny responses to computer generated faces,” Computers in Human Behavior, vol. 25, pp. 695–710, 2009. The research presented in this paper is conducted as part [21] R. McDonnell, M. Breidt, and H. Bulthof ¨ f, “Render me real?: investi- of the project EMPATHIC that has received funding from gating the effect of render style on the perception of animated virtual humans,” ACM Transactions on Graphics, vol. 31, pp. 1–11, 2012. the European Union’s Horizon 2020 research and innovation [22] C. Bartneck, T. Kanda, H. Ishiguro, and N. Hagita, “Is the uncanny programme under grant agreement No 769872. valley an uncanny cliff?” in IEEE Inte. Symp. on Robot and Human interactive Communication, 2007, pp. 368 – 373. [23] T. Burleigh, J. Schoenherr, and G. Lacroix, “Does the uncanny valley REFERENCES exist? an empirical test of the relationship between eeriness and the hu- man likeness of digitally created faces,” Computers in Human Behavior, [1] D. M. Cereghetti, S. Kleanthous, C. Christophorou, C. Tsiourti, vol. 29, pp. 759–771, 2013. C. Wings, and E. Christodoulou, “Virtual partners for seniors: Analysis [24] K. F. MacDorman, “Subjective ratings of robot video clips for human of the users’ preferences and expectations on personality and appear- likeness, familiarity, and eeriness: An exploration of the uncanny valley,” ance.” in AmI (Workshops/Posters), 2015. in ICCS/CogSci-2006 long symposium: Toward social mechanisms of [2] A. Fadhil, “Beyond patient monitoring: Conversational agents role in android science, 2006, pp. 26–29. telemedicine & healthcare support for home-living elderly individuals,” [25] N. Epley, A. Waytz, and J. Cacioppo, “On seeing human: A three-factor CoRR, vol. abs/1803.06000, 2018. theory of anthropomorphism,” Psychological review, vol. 114, pp. 864– [3] J. Olaso, C. Montenegro, A. Vazquez, ´ R. Justo, J. Lozano, N. Dugan, 86, 2007. M. Irvine, C. Pickard, A. Troncone, A. Mtibaa, M. Hmani, M. Korsnes, [26] S. Guthrie, “A cognitive theory of religion,” Current Anthropology - L. Martinussen, S. Escalera, C. Cantarino, ˜ O. Deroo, O. Gordeeva, CURR ANTHROPOL, vol. 21, 1980. J. Tenorio-Laranga, E. Gonzalez-Fraile, ´ and D. Petrovska-Delacretaz, ´ [27] B. Tondu, “Anthropomorphism and service humanoid robots: An “The empathic project: Mid-term achievements,” in Int. Conf. on PErva- ambiguous relationship,” Industrial Robot: An International Journal, sive Technologies Related to Assistive Environments, 2019, p. 629–638. vol. 39, 2012. [4] K. Dautenhahn, “The art of designing socially intelligent agents: Sci- [28] T. Fong, I. Nourbakhsh, and K. Dautenhahn, “A survey of socially ence, fiction, and the human in the loop,” Applied Artificial Intelligence, interactive robots,” Robotics and Autonomous Systems, vol. 42, pp. 143– vol. 12, pp. 573–617, 1998. 166, 2003. [5] P. Kulms and S. Kopp, “More human-likeness, more trust? the effect of [29] L. Ring, D. Utami, and T. Bickmore, “The right agent for the job?” in anthropomorphism on self-reported and behavioral trust in continued Intelligent Virtual Agents, T. Bickmore, S. Marsella, and C. Sidner, Eds. and interdependent human-agent cooperation,” in Proc. of Mensch Cham: Springer International Publishing, 2014, pp. 374–384. und Computer. New York, NY, USA: Association for Computing [30] M. M. Peeters, V. Motti, H. Frijns, S. Mehrotra, T. Akkoc ¸, S. Yengec, Machinery, 2019, p. 31–42. O. C ¸ alik, and M. Neerincx, “Design and development of a physical [6] S. Sturgeon, A. Palmer, J. Blankenburg, and D. Feil-Seifer, “Perception and a virtual embodied conversational agent for social support of older of social intelligence in robots performing false-belief tasks,” in IEEE adults,” in eNTERFACE Summer Workshop on Multimodal Interfaces. Int. Symp. on Robot and Human interactive Communication, Oct 2019, CTIT, 2017, pp. 21–29. pp. 1–7. [31] J.-A. Micoulaud-Franchi, P. Sagaspe, E. De Sevin, S. Bioulac, A. Sauter- [7] B. Duffy, “Anthropomorphism and the social robot,” Robotics and aud, and P. Philip, “Acceptability of embodied conversational agent in a Autonomous Systems, vol. 42, pp. 177–190, 2003. health care context,” in Int. Conf. on Intelligent Virtual Agents. Springer, [8] N. Kushmerick, “Software agents and their bodies,” Minds and Ma- 2016, pp. 416–419. chines, vol. 7, pp. 227–247, 1997. [32] A. Tinwell and R. J. Sloan, “Children’s perception of uncanny human- [9] K. Nowak and F. Biocca, “The effect of the agency and anthropomor- like virtual characters,” Computers in Human Behavior, vol. 36, pp. 286 phism on users’ sense of telepresence, copresence, and social presence in – 296, 2014. virtual environments,” Presence Teleoperators & Virtual Environments, [33] A. Esposito, T. Amorese, M. Cuciniello, A. M. Esposito, A. Troncone, vol. 12, pp. 481–494, 2003. M. I. Torres, S. Schlogl, ¨ and G. Cordasco, “Seniors’ acceptance of [10] E. A. Jentsch, “Zur psychologie des unheimlichen,” Psychiatrisch- virtual humanoid agents,” in Italian Forum of Ambient Assisted Living. neurologische Wochenschrift, vol. 8, no. 22, pp. 195 – 198, 1906. Springer, 2018, pp. 429–443. [11] M. Mori, K. F. MacDorman, and N. Kageki, “The uncanny valley [from [34] R. Yaghoubzadeh and S. Kopp, “Toward a virtual assistant for vulnerable the field],” IEEE Robotics Automation Magazine, vol. 19, no. 2, pp. users: designing careful interaction,” in ACL Workshop on Speech and 98–100, 2012. Multimodal Interaction in Assistive Environments, 2012, pp. 13–17. [12] R. Edu, J. Stasko, and J. Xiao, “Anthropomorphic agents as a user in- [35] C.-C. Ho and K. MacDorman, “Measuring the uncanny valley effect: terface paradigm: Experimental findings and a framework for research,” Refinements to indices for perceived humanness, attractiveness, and GVU Technical Report, vol. 10, 2002. eeriness,” Int. J. of Social Robotics, vol. 9, pp. 129–139, 2017. [13] B. Weiß, I. Wechsung, C. Kuhnel, ¨ and S. Moller ¨ , “Evaluating embodied conversational agents in multimodal interfaces,” Computational Cogni- tive Science, vol. 1, 2015. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Computing Research Repository arXiv (Cornell University)

Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the Uncanny Valley

Download PDF
6 pages

Loading...

Page 2

Loading...

Page 3

Loading...

Page 4

Loading...

Page 5

Loading...

Page 6

 
/lp/arxiv-cornell-university/agent-vs-avatar-comparing-embodied-conversational-agents-concerning-c4AtuK5TTn

References (36)

eISSN
ARCH-3344
DOI
10.1109/ICHMS49158.2020.9209539
Publisher site
See Article on Publisher Site

Abstract

Agent vs. Avatar: Comparing Embodied Conversational Agents Concerning Characteristics of the Uncanny Valley Markus Thaler Stephan Schlogl ¨ Aleksander Groth Management, Communication & IT Management, Communication & IT Management, Communication & IT MCI – The Entrepreneurial School MCI – The Entrepreneurial School MCI – The Entrepreneurial School Innsbruck, Austria Innsbruck, Austria Innsbruck, Austria mark.thaler@mci4me.at stephan.schloegl@mci.edu aleksander.groth@mci.edu Abstract—Visual appearance is an important aspect influenc- where a too realistic appearance is found to trigger feelings ing the perception and consequent acceptance of Embodied Con- of eeriness in the human interlocutor – a phenomenon that versational Agents (ECA). To this end, the Uncanny Valley theory th was already described in the early 20 century by Jentsch, contradicts the common assumption that increased humanization observing interactions with automatic dolls [10]. The heron of characters leads to better acceptance. Rather, it shows that built Uncanny Valley effect demonstrates that the affinity anthropomorphic behavior may trigger feelings of eeriness and rejection in people. The work presented in this paper explores with a given object is based on a non-linear relationship whether four different autonomous ECAs, specifically build for with the human-likeness of the respective object [11]. Our a European research project, are affected by this effect, and how study examines to which extent this effect can be found they compare to two slightly more realistically looking human- with the Embodied Conversational Agents (ECA) used in the controlled, i.e. face-tracked, ECAs with respect to perceived European research project EMPATHIC . In particular, our humanness, eeriness, and attractiveness. Short videos of the ECAs in combination with a validated questionnaire were used to goal was to investigate the following research question: investigate potential differences. Results support existing theories highlighting that increased perceived humanness correlates with How do human-driven, i.e. face-tracked, ECAs (so-called increased perceived eeriness. Furthermore, it was found, that avatars) compare to autonomous ECAs with respect to their neither the gender of survey participants, their age, nor the perceived humanness, eeriness and attractiveness? sex of the ECA influences this effect, and that female ECAs are perceived to be significantly more attractive than their male II. RELATED WORK counterparts. Index Terms—Embodied Conversational Agents, Avatars, Un- Although recent years have seen great progress in the devel- canny Valley, Humanness, Eeriness, Attractiveness opment of speech and language-based user interfaces, people still feel uneasy when ‘talking’ to an artificial entity [12], [13]. I. INTRODUCTION One reason for this may be rooted in the fact, that a large For decades, research and industry have been struggling to part of human communication lies outside the scope of pure build human-like computing systems, i.e. digital assistants, language. ECAs aim to ameliorate this shortcoming through an that can be operated via natural interaction modalities (e.g. additional emotional interaction channel. That is, they not only natural language, gestures, mimics, etc.) and are capable support speech-based interaction [13], [14], but also feature of understanding and expressing emotions. Due to past some sort of visual and/or physical appearance [15], which al- developments in artificial intelligence and natural language lows for the generation of facial expressions. Consequently, the processing, and the increasing uptake of digital assistants human interlocutors’ abilities to predict an agent’s behaviour in the consumer sector, this gap between science fiction improve [16]. In addition, the embodiment lets an agent appear and reality is, however, gradually closing. With respect more aesthetic and life-like [4], [17] as well as potentially to application areas for these types of intelligent systems, more intelligent [8]. Yet, improvement stops once the agent research is often looking at the social and health care crosses the so-called Uncanny Valley (cf. Fig. 1). context, where people may benefit from support in various A. The Uncanny Valley quotidian tasks [1]–[3]. Here, the question as to which social characteristics a digital agent should or should not exhibit in According to Mori [11], the Uncanny Valley (UV) shows these settings, is continuously being researched [4]–[6]. One the affinity a human observer has towards an object and how of these issues concerns the right embodiment of respective such changes when the human likeness of the object increases. systems, and how such may influence their credibility [7]–[9]. Humanization, in particular, may pose unwanted side-effects http://www.empathic-project.eu/ arXiv:2104.11043v1 [cs.HC] 22 Apr 2021 B. Anthropomorphism The anthropomorphism phenomenon is one crucial aspect related to the UV theory, stating that human characteristics, motivations, intentions and emotions may be attributed to non- human entities, even if these have only slightly human-like traits [25]. Duffy [7, p. 180] describes this as “attributing cog- nitive or emotional states to something based on observation in order to rationalise an entity’s behaviour in a given social environment”. In other words, anthropomorphism is consid- ered a psychological process helping humans rationalize ob- servations. Guthrie [26] describes the respective procedure in Fig. 1. The Uncanny Valley according to Mori [11] four consecutive steps: (1) one is confronted with an unknown object which initially triggers shallow conclusions; (2) this initial conclusions are then combined with and expanded upon past experiences; (3) the most probable template describing Ho & MacDorman [18] describe this as a graph illustrating a the object is selected; and (4) since the ‘human’ template non-linear relationship between the human resemblance of an is most commonly chosen, the observed characteristics are anthropomorphic entity and a viewer’s emotional response. As compared to those of humans. As this process is based on can be seen in Fig. 1, Mori chose industrial robots as a starting non-rational thinking, humanization can be perceived nega- point, since they exhibit little resemblance to humans. At the tively [27]. Contrary to this, in Human-Computer Interaction opposite side of the valley one then finds for example realistic research and design, anthropomorphism is often considered hand prostheses, as they are rather difficult to distinguish from a design guideline that helps build appealing user interfaces real human hands. According to Seyama & Nagayama [19], (UI), and aims at triggering rather pleasant experiences in the UV effect can be shown with both physical and virtual users [4]. While humans often anthropomorphise, even in objects. In an experiment with computer-generated characters cases where an interface exhibits hardly any human traits, MacDorman and colleagues, for example, found that faces it has been shown that the context and given task setting with photo-realistic textures and a high degree of detail tend to influence the perception and consequent acceptance of this trigger eeriness. Characters with a lower photo-realism level human-likeness [28]. In social tasks, for example, a human- (i.e., 75%), however, were deemed the least uncanny by human like agent seems acceptable, whereas in other settings an viewers [20]. Hence, it may be argued that the UV theory anthropomorphic appearance may be less expected, and con- contradicts the view that a robot or embodied agent should sequently trigger feelings of uneasiness [29]. resemble a human being as close as possible [4]. The second part of the UV theory deals with motion. C. ECAs in Social and Health Care Settings According to Mori [11], the amplitude of the curve increases ECAs may help simplify human-technology interaction. when the object is in motion. In an experiment with avatars For seniors, in particular, accessibility to technology can be (i.e. human-controlled agents), it was found that motion influ- improved [3], [30]. One area that sees an ongoing uptake of ences familiarity and perceived attractiveness of the agent [21]. these types of intelligent systems is the social and health care Although, it has to be underlined that both familiarity and sector, where ECAs are increasingly used to support everyday perceived attractiveness are subject to personal taste. For tasks of people [31]. However, as already outlined above, the example, it was shown that even very human-like characters visual appearance of these ECAs poses certain challenges with may be deemed repelling [19]. Similarly, Jentsch notes in respect to their human-likeness and the consequent effects on his definition of uncannyness, that a realistic technological people’s feelings of eeriness. Thus, the goal of our work was to stimulus can trigger a feeling of eeriness in people [10]. There explore how ECAs specifically created for health related tasks, are several theories where this sensation of eeriness might i.e. the ECAs built for the EMPATHIC project, are perceived come from [19]. Bartneck et al. attribute it to the framing by the general population. theory [22], which states that when observing new things, III. M ETHODOLOGY humans refer to past experiences, so-called frames. This gives rise to expectations that may not always be met. With human- The goal was to evaluate six different ECAs (cf. Figures 2 like robots, for example, the ‘human frame’ is selected, but & 3). Four autonomous agents, which were created for the when seeing the robot’s mechanical movements, which may EMPATHIC project, i.e. Lena, Alice, Christian and Adam, resemble those of a sick or injured person, it is often the case and two human-driven/face-tracked agents (so-called avatars), that a feeling of discomfort arises. Another trigger for this i.e. Sophie and Michael, which were built separately, so as eeriness may be found in the skin discoloration of artificial to serve as a control group (note: for these two agents we characters, which reminds people of death and thus may evoke 2 used the Reallusion Software Suite , which led to slightly respective fears [14], [23], [24]. Here one might also draw a connection to the psychological process of anthropomorphism. https://www.reallusion.com/iclone/ perceived attractiveness dependent on the sex of the ECA (cf. Esposito et al. [33]). H5a: There is a significant positive correlation between a participant’s age and the perceived eeriness of ECAs (cf. Yaghoubzadeh et al. [34]). H5b: There is a significant negative correlation between a participant’s age and the perceived attractiveness of ECAs (cf. Yaghoubzadeh et al. [34]). H6: Human-driven ECAs (i.e., avatars) exhibit a higher level of perceived eeriness than autonomous ECAs. In order to evaluate these hypotheses and consequently the extent to which the tested ECAs may fall into the UV, we exposed participants to an eight-second video of an agent and subsequently asked them to rate the agent’s perceived level of humanness, eeriness and attractiveness according to the following 21 UV semantic differential effect scales proposed Fig. 2. Autonomous ECAs built for the EMPATHIC project by MacDorman & Ho [35] (note: in brackets we provide the German translation of the terms the way they were used in the questionnaire): Humanness: 7-point semantic differentials on – inanimate $ living (ge: unbelebt/lebendig) – synthetic $ real (ge: synthetisch/echt) – mechanical movement $ biological movement (ge: mech. Bewegungen/nat. Bewegungen) – human-made $ humanlike (ge: von Menschen gemacht/menschenahnlich) – without definite lifespan $ mortal Fig. 3. Human-driven ECAs (built using the Reallusion Software Suite) (ge: ohne definitive Lebensdauer/sterblich) – artificial $ natural (ge: kunstlich/nat ¨ urlich) ¨ Eeriness: 7-point semantic differentials on more realistic faces and facial expressions than what was – dull $ freaky (ge: eintonig/ausgef ¨ allen) achieved by the project ECAs). All six ECAs exhibited similar – predictable $ eerie (ge: abschatzbar/unheimlich) ¨ characteristics with respect to Ring et al.’s classification of – plain $ weird (ge: schlicht/sonderbar) anthropomorphic UIs [29]. That is, all were human(-like), – ordinary $ supernatural mimicking the same ethnicity and similar age and wearing (ge: gewohnlich/außernat ¨ urlich) ¨ similar clothing. Only the level of realism distinguished the – boring $ shocking (ge: langweilig/schockierend) project ECAs (i.e., Lena, Alice, Christian and Adam) from – uninspiring $ spine-tingling the slightly more realistic Reallusion ECAs (i.e., Sophie and Michael). In accordance with the above discussed literature (ge: uninspirierend/elektrisierend) we thus formed the following hypotheses to be tested: – predictable$ thrilling (ge: vorhersehbar/mitreißend) – bland $ uncanny (ge: nichtssagend/untypisch) H1: Human-driven ECAs (i.e., avatars) exhibit a higher – unemotional $ hair-raising level of perceived humanness than autonomous ECAs (cf. (ge: emotionslos/furchterregend) MacDorman et al. [20]). – familiar $ uncanny (ge: vertraut/unheimlich) H2: The perceived level of humanness has a significant Attractiveness: 7-point semantic differentials on influence on the perceived level of eeriness (cf. McDon- – ugly $ beautiful (ge: hasslich/sch ¨ on) ¨ nell et al. [21]). – repulsive $ agreeable (ge: abstoßend/ansprechend) H3a: There is a significant difference between genders – crude $ stylish (ge: geschmackslos/stylisch) when it comes to the perceived level of eeriness (cf. – messy $ sleek (ge: unordentlich/gepflegt) Tinwell & Sloan [32]). – unattractive $ attractive (ge: unattraktiv/attraktiv) H3b: There is a significant difference between genders when it comes to the perceived level of attractiveness This procedure was repeated for each of the six ECAs. (cf. Tinwell & Sloan [32]). Finally, participants were asked to provide some additional H4a: There is a significant difference with respect to demographic information. We particularly targeted German- perceived eeriness dependent on the sex of the ECA (cf. speaking participants so as to prevent any potential cultural Esposito et al. [33]). bias. To this end, all scales and questions were translated. A H4b: There is a significant difference with respect to German-speaking lecturer of English as well as a bilingual TABLE I R ELIABILITY OF S CALES Stimulus Humanness ( ) Attractiveness ( ) Eeriness ( ) Lena 0.871 0.896 0.827 Adam 0.925 0.896 0.856 Sophie 0.937 0.932 0.864 Christian 0.956 0.910 0.878 Alice 0.948 0.921 0.878 Michael 0.947 0.951 0.894 Overall 0.957 0.926 0.950 Fig. 4. Mean Values of Perceived Humanness for Each of the Shown ECAs colleague were consulted to improve translation quality. In TABLE II addition, all scales were tested for reliability before further ANOVA - POST-HOC S CHEFFE evaluations began. Participation in the study was voluntary Stimuli Humanness (p) Eeriness (p) (note: three EUR 10,- Amazon Gift Vouchers were raffled Alice ! Adam 0.0015** 0.8295 off among the participants). Furthermore, questionnaire and Christian ! Adam 0.9994 0.9988 respective procedure were evaluated by MCI’s research ethics Lena ! Adam 0.9995 0.9976 Michael ! Adam 0.0000*** 0.0000*** group for accordance with ethical guidelines concerning re- Sophie ! Adam 0.0000*** 0.0903 search with human participation. Christian ! Alice 0.0066** 0.5834 Lena ! Alice 0.0064** 0.5441 IV. R ESULTS Michael ! Alice 0.3798 0.0068** Sophie ! Alice 0.5038 0.7548 A total of n = 215 participants (150 females) completed Lena ! Christian 1.0000 1.0000 the above described procedure, approx. 50% (i.e. 108) of Michael ! Christian 0.0000*** 0.0000*** whom were students (predominantly business and information Sophie ! Christian 0.0000*** 0.0266* Michael ! Lena 0.0000*** 0.0000*** systems but also other fields such as pharmacy and physics). Sophie ! Lena 0.0000*** 0.0219* The overall age distribution (74%  30 years; 21% 31 Sophie ! Michael 1.0000 0.3367 60 years; 5% > 60 years) shows a strong right-bound skewness of 1.766 (Mean = 31:04; Median = 25). Scale reliability for humanness, eeriness and attractiveness were first in perceived humanness (p = 0:014) attributed to the type of tested individually for each stimulus. Then each dimension control, i.e. autonomous vs. human-controlled. An ANOVA was analyzed in its entirety using Cronbach’s . Table I shows including the Scheffe ´ post-hoc test generally confirmed this that the overall scale reliability was excellent ( = 0:95) significantly higher level of perceived humanness with the and when focusing on internal stimuli only, it was still good Reallusion ECAs, yet further showed that also Alice (the most ( > 0:70) [35]. Given this reliability of scales we were human-like project ECA) scored significantly higher than the able to proceed, evaluating and consequently comparing the other project ECAs (cf. Table II). All in all, however, we may six ECAs with respect to their perceived humanness, eeriness, argue that our data supports H1. and attractiveness. Respective results are summarized in the following sub-sections. B. Eeriness A. Humanness In order to explore the UV effect, ECAs were ordered Fig. 4 shows the mean values of the six ECAs with respect according to their human similarity scores and then evalu- to humanness (note: we calculated the mean over all the ated with respect to their perceived level of eeriness. Fig. 5 scales concerning humanness described in Section III where shows that after initial improvements eeriness starts increasing values close to 1 signify low humanness and values close to with increasing humanness. That is, the project ECAs Adam 7 signify high humanness). To this end, the project ECAs (x  = 3:19; SD = 0:94), Lena (x  = 3:14; SD = 0:80) Adam (x  = 2:58; SD = 1:22), Lena (x  = 2:63; SD = 1:09) and Christian (x  = 3:15; SD = 0:95) had lower eeriness and Christian (x  = 2:63; SD = 1:36) were perceived quite scores than Alice (x  = 3:32; SD = 0:89). Similarly, the similarly. Alice, however, was perceived as more human-like Reallusion ECAs Sophie (x  = 3:46; SD = 0:90) and Michael (x  = 3:15; SD = 1:39). The two Reallusion ECAs Sophie (x  = 3:67; SD = 1:01) scored rather negatively, with Michael (x  = 3:41; SD = 1:38) and Michael (x  = 3:44; SD = 1:48) moving even further away from Sophie and again showing were also perceived similarly and more human-like than any the greatest variability in people’s perception of eeriness. A of the project ECAs. Yet, Michael’s ratings showed the highest T-test for independent samples showed that this perception variability. Although, human similarity exhibited generally a was independent of people’s gender (F-test for gender variance higher standard deviation than the other dimensions. equality: p = 0:752), for which H3a had to be rejected (p = A T-test for combined samples between the most human- 0:655). Similarly with respect to the agents’ sex no significant like project ECA (i.e., Alice) and the least human-like Re- difference in perceived eeriness was found (p = 0:411). Hence, allusion ECA (i.e., Sophie) points to a significant difference also H4a was rejected. Fig. 5. Mean Values of Perceived Eeriness for Each of the Shown ECAs Fig. 6. Mean Values of Perceived Attractiveness for Each of the Shown ECAs Looking at a connection between humanness and eeriness, a Finally, investigating a potential negative relation between Pearson correlation analysis points to a medium to strong pos- the participants’ age and the perceived attractiveness of ECAs, itive correlation between those two variables (r = 0:573; p = no significant correlation was found (r = 0:095; p = 0:164), 0:000). The subsequently performed linear regression shows for which also H5b had to be rejected. that perceived humanness is able to explain 32.9% of the vari- ance in perceived eeriness (R = 0:329; Beta = 0:573; p = V. S UMMARY, L IMITATIONS AND OUTLOOK 0:000). Hence, H2 is supported by the data. The goal of our study was to examine six different ECAs With respect to H5a, we explored a potential positive concerning characteristics of the UV. According to the clas- relation between a participant’s age and the ECA’s perceived sification by Ring and colleagues [29] we expected a more eeriness. The performed Pearson correlation analysis, however, realistic rendering style to perform better than a less realistic shows a weak but significant negative connection between rendering style. Consequently, we evaluated the ECAs with re- those two variables (r = 0:150; p = 0:014). Consequently, spect to their perceived humanness, attractiveness and eeriness. H5a had to be rejected. The results of our evaluation do not fully confirm our expecta- Finally, a T-test for combined samples pointed to a sig- tions, yet they are in line with previous work. That is, we found nificant difference between the most eerie project ECA (i.e., a positive correlation between an ECA’s human similarity and Alice) and the least eerie Reallusion ECA (i.e., Sophie), its perceived eeriness. Analogous to the UV theory, this means supporting the assumption outlined by H6 (p = 0:024). The that with increasing realism, the viewer’s affinity with an ECA subsequently performed ANOVA and Scheffe ´ post-hoc test decreases. Age-related connections could only be confirmed to confirmed this significantly higher level of perceived eeriness a limited extent, showing a slight negative correlation with the for Michael (i.e., significantly higher eeriness scores than all participant’s eerieness ratings. The participant’s gender had project ECAs) and partly for Sophie (i.e., significantly higher no effects on the ratings, the sex of the stimuli, however, eeriness scores than Christian and Lena). Overall, we thus influenced their perceived attractiveness, with female agents may argue that our data supports H6 in that the perceived being perceived significantly more attractive than male ones. eeriness of the two Reallusion ECAs (i.e., the avatars) was Although these results generally confirm previous work, we rated significantly higher than the perceived eeriness of the would like to point to a number of limitations of our study. four project ECAs (cf. Table II). First, while our human-driven ECAs were indeed perceived C. Attractiveness more human-like than the autonomous ECAs built for the EMPATHIC project, their level of humanness was still rather Attractiveness was generally rated higher than the other low (i.e., x  = 3:41 and x  = 3:44 respectively). Such may two dimensions. Lena (x  = 4:61; SD = 1:01) and Sophie have had an affect on the significance of results, in particular (x  = 4:40; SD = 1:21) were rated the most attractive, with respect to the origin of perceived eeriness. Second, the Adam the least attractive (x  = 3:43; SD = 1:07), and Alice (x  = 4:16; SD = 1:17), Michael (x  = 4:07; SD = 1:35) and used questionnaire was rather long and monotonous, repeating Christian (x  = 3:96; SD = 1:08) lay somewhere in between. the same sections (i.e. MacDorman & Ho’s 21 semantic Also here the participants agreed the least with the evaluation differential effect scales) six times, which may have influenced of Michael. In Fig. 6 the ECAs were ranked according to their people’s responses. Third, the negative correlation we found perceived humanness. Looking at the graph, it is noticeable with respect to participants’ age and their provided eeriness that the female agents were consistently classified as more ratings was very weak (cf. H5a). Finally, a more homogeneous attractive. A T-test on the mean values of the ECA’s sex sample may have potentially provided better insights. That is, supports this assumption, showing a significant difference in although in general our data did not point to any irregularities, terms of perceived attractiveness (p = 0:000). Consequently, the demographic distribution was not ideal. The majority of H4b is supported by the data. From a participants’ point respondents were students younger than 30 years, and more of view, however, no gender differences with respect to the than two-thirds were female. perceived attractiveness of ECAs was found (p = 0:278). In conclusion, this study has confirmed previous findings Hence H3b had to be rejected. with respect to the UV theory; i.e. the more human-like an ECA appears the greater its produced feeling of eeriness. An [14] L. Ciechanowski, A. Przegalinska, M. Magnuski, and P. A. Gloor, “In the shades of the uncanny valley: An experimental study of human-chatbot additional, not negligible result of our study, is the translation interaction,” Future Generation Comp. Syst., vol. 92, pp. 539–548, 2018. and validation of MacDorman & Ho’s UV scales into German, [15] S. Franklin and A. Graesser, “Is it an agent, or just a program?: A tax- which may allow for additional future studies in German- onomy for autonomous agents,” in Intelligent Agents III Agent Theories, Architectures, and Languages, J. P. Muller ¨ , M. J. Wooldridge, and N. R. speaking countries. Those studies may, for example, focus Jennings, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 1997, on the level of realism that is accepted with ECAs. To this pp. 21–35. end an interesting question to investigate may be whether [16] T. Koda and P. Maes, “Agents with faces: the effect of personification,” IEEE Int. Symp. on Robot and Human interactive Communication, pp. ECAs could be made so realistic and credible that they would 189 – 194, 1996. ‘climb up’ the other end of the UV. Also, it may be worth [17] J. Cassell, “Embodied conversational agents: Representation and intel- exploring whether female ECAs are consistently rated more ligence in user interfaces.” AI Magazine, vol. 22, pp. 67–84, 2001. [18] C.-C. Ho and K. MacDorman, “Revisiting the uncanny valley hypothe- attractive than male ones. Finally, concerns regarding privacy sis: Developing and validating an alternative to the godspeed indices,” and ethics connected to the use of ECAs would require Computers in Human Behavior, vol. 26, pp. 1508–1518, 2010. additional research. In particular, if their potential application [19] J. Seyama and R. Nagayama, “The uncanny valley: Effect of realism on the impression of artificial human faces,” Presence, vol. 16, pp. 337–351, domain is focusing on social and health care settings. [20] K. MacDorman, R. Green, C.-C. Ho, and C. Koch, “Too real for ACKNOWLEDGMENT comfort? uncanny responses to computer generated faces,” Computers in Human Behavior, vol. 25, pp. 695–710, 2009. The research presented in this paper is conducted as part [21] R. McDonnell, M. Breidt, and H. Bulthof ¨ f, “Render me real?: investi- of the project EMPATHIC that has received funding from gating the effect of render style on the perception of animated virtual humans,” ACM Transactions on Graphics, vol. 31, pp. 1–11, 2012. the European Union’s Horizon 2020 research and innovation [22] C. Bartneck, T. Kanda, H. Ishiguro, and N. Hagita, “Is the uncanny programme under grant agreement No 769872. valley an uncanny cliff?” in IEEE Inte. Symp. on Robot and Human interactive Communication, 2007, pp. 368 – 373. [23] T. Burleigh, J. Schoenherr, and G. Lacroix, “Does the uncanny valley REFERENCES exist? an empirical test of the relationship between eeriness and the hu- man likeness of digitally created faces,” Computers in Human Behavior, [1] D. M. Cereghetti, S. Kleanthous, C. Christophorou, C. Tsiourti, vol. 29, pp. 759–771, 2013. C. Wings, and E. Christodoulou, “Virtual partners for seniors: Analysis [24] K. F. MacDorman, “Subjective ratings of robot video clips for human of the users’ preferences and expectations on personality and appear- likeness, familiarity, and eeriness: An exploration of the uncanny valley,” ance.” in AmI (Workshops/Posters), 2015. in ICCS/CogSci-2006 long symposium: Toward social mechanisms of [2] A. Fadhil, “Beyond patient monitoring: Conversational agents role in android science, 2006, pp. 26–29. telemedicine & healthcare support for home-living elderly individuals,” [25] N. Epley, A. Waytz, and J. Cacioppo, “On seeing human: A three-factor CoRR, vol. abs/1803.06000, 2018. theory of anthropomorphism,” Psychological review, vol. 114, pp. 864– [3] J. Olaso, C. Montenegro, A. Vazquez, ´ R. Justo, J. Lozano, N. Dugan, 86, 2007. M. Irvine, C. Pickard, A. Troncone, A. Mtibaa, M. Hmani, M. Korsnes, [26] S. Guthrie, “A cognitive theory of religion,” Current Anthropology - L. Martinussen, S. Escalera, C. Cantarino, ˜ O. Deroo, O. Gordeeva, CURR ANTHROPOL, vol. 21, 1980. J. Tenorio-Laranga, E. Gonzalez-Fraile, ´ and D. Petrovska-Delacretaz, ´ [27] B. Tondu, “Anthropomorphism and service humanoid robots: An “The empathic project: Mid-term achievements,” in Int. Conf. on PErva- ambiguous relationship,” Industrial Robot: An International Journal, sive Technologies Related to Assistive Environments, 2019, p. 629–638. vol. 39, 2012. [4] K. Dautenhahn, “The art of designing socially intelligent agents: Sci- [28] T. Fong, I. Nourbakhsh, and K. Dautenhahn, “A survey of socially ence, fiction, and the human in the loop,” Applied Artificial Intelligence, interactive robots,” Robotics and Autonomous Systems, vol. 42, pp. 143– vol. 12, pp. 573–617, 1998. 166, 2003. [5] P. Kulms and S. Kopp, “More human-likeness, more trust? the effect of [29] L. Ring, D. Utami, and T. Bickmore, “The right agent for the job?” in anthropomorphism on self-reported and behavioral trust in continued Intelligent Virtual Agents, T. Bickmore, S. Marsella, and C. Sidner, Eds. and interdependent human-agent cooperation,” in Proc. of Mensch Cham: Springer International Publishing, 2014, pp. 374–384. und Computer. New York, NY, USA: Association for Computing [30] M. M. Peeters, V. Motti, H. Frijns, S. Mehrotra, T. Akkoc ¸, S. Yengec, Machinery, 2019, p. 31–42. O. C ¸ alik, and M. Neerincx, “Design and development of a physical [6] S. Sturgeon, A. Palmer, J. Blankenburg, and D. Feil-Seifer, “Perception and a virtual embodied conversational agent for social support of older of social intelligence in robots performing false-belief tasks,” in IEEE adults,” in eNTERFACE Summer Workshop on Multimodal Interfaces. Int. Symp. on Robot and Human interactive Communication, Oct 2019, CTIT, 2017, pp. 21–29. pp. 1–7. [31] J.-A. Micoulaud-Franchi, P. Sagaspe, E. De Sevin, S. Bioulac, A. Sauter- [7] B. Duffy, “Anthropomorphism and the social robot,” Robotics and aud, and P. Philip, “Acceptability of embodied conversational agent in a Autonomous Systems, vol. 42, pp. 177–190, 2003. health care context,” in Int. Conf. on Intelligent Virtual Agents. Springer, [8] N. Kushmerick, “Software agents and their bodies,” Minds and Ma- 2016, pp. 416–419. chines, vol. 7, pp. 227–247, 1997. [32] A. Tinwell and R. J. Sloan, “Children’s perception of uncanny human- [9] K. Nowak and F. Biocca, “The effect of the agency and anthropomor- like virtual characters,” Computers in Human Behavior, vol. 36, pp. 286 phism on users’ sense of telepresence, copresence, and social presence in – 296, 2014. virtual environments,” Presence Teleoperators & Virtual Environments, [33] A. Esposito, T. Amorese, M. Cuciniello, A. M. Esposito, A. Troncone, vol. 12, pp. 481–494, 2003. M. I. Torres, S. Schlogl, ¨ and G. Cordasco, “Seniors’ acceptance of [10] E. A. Jentsch, “Zur psychologie des unheimlichen,” Psychiatrisch- virtual humanoid agents,” in Italian Forum of Ambient Assisted Living. neurologische Wochenschrift, vol. 8, no. 22, pp. 195 – 198, 1906. Springer, 2018, pp. 429–443. [11] M. Mori, K. F. MacDorman, and N. Kageki, “The uncanny valley [from [34] R. Yaghoubzadeh and S. Kopp, “Toward a virtual assistant for vulnerable the field],” IEEE Robotics Automation Magazine, vol. 19, no. 2, pp. users: designing careful interaction,” in ACL Workshop on Speech and 98–100, 2012. Multimodal Interaction in Assistive Environments, 2012, pp. 13–17. [12] R. Edu, J. Stasko, and J. Xiao, “Anthropomorphic agents as a user in- [35] C.-C. Ho and K. MacDorman, “Measuring the uncanny valley effect: terface paradigm: Experimental findings and a framework for research,” Refinements to indices for perceived humanness, attractiveness, and GVU Technical Report, vol. 10, 2002. eeriness,” Int. J. of Social Robotics, vol. 9, pp. 129–139, 2017. [13] B. Weiß, I. Wechsung, C. Kuhnel, ¨ and S. Moller ¨ , “Evaluating embodied conversational agents in multimodal interfaces,” Computational Cogni- tive Science, vol. 1, 2015.

Journal

Computing Research Repository – arXiv (Cornell University)

Published: Apr 22, 2021

There are no references for this article.