In dialogue with an avatar, language behavior is identical to dialogue with a human partner

Abstract

The use of virtual reality (VR) as a methodological tool is becoming increasingly popular in behavioral research as its flexibility allows for a wide range of applications. This new method has not been as widely accepted in the field of psycholinguistics, however, possibly due to the assumption that language processing during human-computer interactions does not accurately reflect human-human interactions. Yet at the same time there is a growing need to study human-human language interactions in a tightly controlled context, which has not been possible using existing methods. VR, however, offers experimental control over parameters that cannot be (as finely) controlled in the real world. As such, in this study we aim to show that human-computer language interaction is comparable to human-human language interaction in virtual reality. In the current study we compare participants' language behavior in a syntactic priming task with human versus computer partners: we used a human partner, a human-like avatar with human-like facial expressions and verbal behavior, and a computer-like avatar which had this humanness removed. As predicted, our study shows comparable priming effects between the human and human-like avatar suggesting that participants attributed human-like agency to the human-like avatar. Indeed, when interacting with the computer-like avatar, the priming effect was significantly decreased. This suggests that when interacting with a human-like avatar, sentence processing is comparable to interacting with a human partner. Our study therefore shows that VR is a valid platform for conducting language research and studying dialogue interactions in an ecologically valid manner.

Keywords: Human-computer interaction; Language; Structural priming; Syntactic processing; Virtual reality.

Fig. 1

Avatar. The exterior of the avatar was identical for both avatar partners

Fig. 2

Set-up. (A) The experimental set-up from the view of the participant. The only difference is that in the virtual environment (VE) the cards were presented at the top of the divider, whereas in the Human block, the cards were laid out on the table. (B) The participant card (left) and confederate card (right). The participant card only showed the neutral verb associated with the photo, whereas the confederate card had a complete sentence written underneath. Here “to kiss” and “The man kisses the woman”

Fig. 3

Proportion of passive responses per prime type for Experiment 1. As predicted, there are no significant differences in syntactic priming effects between the human and the avatar block. Passive production increased by 11.8 % for the human block and by 12.3 % for the avatar block following a passive prime compared to the baseline condition. In line with previous research, there were no priming effects for actives. Error bars represent standard error

Fig. 4

Cumulativity of passive responses for Experiment 1. The proportion of passive responses produced increases for both partner types over the course of the block. Mixed models show that there is a significant difference between the probability of producing a passive response between human and avatar blocks (p = .012). This is most likely due to the lower starting point of the avatar partner. The learning curve (between trial 0 and 75) is equally steep for other partner types

Fig. 5

Main effect of liking on Passive Production. The more likeable the participant rated their interlocutor, the more passive responses the participant produced with that interlocutor (p = .003). This effect was not significantly different between human and avatar partner

Fig. 6

Effect of Dominance on Passive Priming. With increasing self-ratings of Dominance in Conflict, participants produced less passive responses compared to participants who rated themselves as less dominant in conflict situations. The model stated that there is a significant difference between how responses are effected based on their prime time (active vs. passive; p = .040); however, upon closer observation this effect may be influenced by the variability. Error clouds represent standard error

Fig. 7

Humanness and familiarity ratings of the two avatar types. Ratings were given immediately after the encounter with the avatar, although participants were able to change their answer after they had been exposed to both. Error bars represent standard error. The computer-like avatar was rated as significantly less human (p < .0001)

Fig. 8

Proportion of passive responses per prime type for Experiment 2. There are significant differences in syntactic priming effects between the two avatar types (p = .033). Passive production increased by 9.5 % with the human-like avatar and only 3.7 % with the computer-like avatar following a passive prime compared to the baseline condition, which confirmed our prediction that participants primed less with the computer-like avatar as it is less human-like. In line with previous research, there were no priming effects for actives

Fig. 9

Cumulativity of passive responses for Experiment 2. The proportion of passive responses produced increases for both partner types over the course of the block. Mixed models show that there is a significant difference between the two avatar types (p = .0007)

Fig. 10

Effect of dominance on passive production per partner type. (A) The effects for Experiment 2. As self-ratings of dominant behavior in a conflict situation increase, the proportion of passive responses produced also increases. Curiously, B shows the opposite trend for Experiment 1. The human-like avatar is identical in both experiments, showing that this trend is most likely caused by the group make-up being different between experiments. This highlights the sensitivity of social factors to individual differences

Fig. 11

Priming magnitude per partner type. As priming with the human-like avatar was not significantly different between experiments (p = .85), the data are collapsed across experiments. Participants primed comparably with human and human-like avatar partners, but significantly less with the computer-like avatar (p = .03). As the only difference between the avatars was the humanness rating, the results suggest that the high priming magnitude of the human-like avatar is due to its perceived humanness