Converting Existing Vocabulary Tests into Robotic Programs

Background: Recent studies (e.g., Walsh Matthews et al. 2016; Kim et al. 2015; Ismail et al., 2012) show increased social communicative behaviour by children with Autism Spectrum Disorder (ASD) following human robot interaction (HRI). To further these test effects, we designed a study that would substitute human-conducted vocabulary testing with HRI by re-creating Tager-Flusberg’s (1985) experiment to examine the nature of substantive lexical semantic representations in children with ASD. The HRI programming allowed for the examination of basic level and superordinate semantic categories might aid in cognitive representational tasks and that longer, more intensive interactive robotic programming would promote the production of more accurate responding by minimising the observers’ impact during in-session observations.

Objectives: To program novel robotic behaviours on the small humanoid NAO to mimic the Peabody Vocabulary Test (PVT) via randomized Robotic Vocabulary Tests (RVT.) The test is to be administered with children previously diagnosed with ASD. Participants would have an opportunity to interact with NAO to see which of the dances/songs would promote greater interactivity between participant and robot. From our first observations, rewards were generated at different intervals and of different lengths to further evaluate the role of interaction and its relevant effect on responses.

Methods: The study is grounded in Tager-Flusberg’s (1985) experiment 1 to assess autistic children’s understanding of words “for basic and superordinate level categories” (1169). Replacing the PVT with a programmed RVT version, stimuli (images representing two overall categories: superordinate including tools and food and basic including boat and birds ) were transferred into accessible image cards recognised by the robot. Within each category there are three possible stimulus types: match (central members of a category), false lure (related nonmembers), and unrelated objects (unrelated nonmembers). The robotic script runs four randomized RVTs which include two short versions with 5-second robotic dances and songs and two longer versions that have multiple and randomized robotic diversion programming. Select ASD participants (aged 4-9), interacted with NAO once per week for three consecutive weeks.

Results: The length of HRI did not have a significant impact on the participant’s overall accuracy performance at both time points (p = 0.87 and p = 0.54, respectively). However, when analyzing the performance based on category level, where we would expect to find differences, we see differential performance. Based on our preliminary analyses, (see Figure 1), participants perform better on basic level cards when the interaction is long compared to short, and a slight increase in performance for superordinate level cards when the interaction is long compared to short.

Conclusions: Findings from Tager-Flusberg’s original experiment found children likely had inflexible categorical semantic organization. Inconsistent with this, we see that children show improvements in their semantic organization when the robot’s behaviours are more interactive. Further analyses will be needed to confirm this trend, however, early findings supports our hypotheses driven by Idealised Cognitive Models (Lakoff, 1987), such that HRI would increase the potential for new conceptual categories, or at least invite opportunities to engage with them.