Socially Interactive Agents as Peers

Bringing together multimodal and multilevel approaches to study the emergence of social bonds between children and improve social AI

“…In what follows we therefore propose a novel study and concomitant methodology to address some of the questions raised by prior work, and to adduce evidence for the claims laid out above. We have adapted an ecologically valid interactive task used in our prior work (Finkelstein, 2018;Cassell, 2022), attractive to children across middle childhood, that can be carried out by one child (solo phase), and by dyads (collaborative phase), while attached to fNIRS apparatus. The task has variations that allow children to engage several times over a period of weeks, each time with a solo phase and a collaborative phase, and calling on slightly different skills each time, but demanding the same amount of effort.…”

“…The age of the participants indicates that they will have a relatively short attention span, meaning that the task must be able to be completed relatively rapidly. On the other hand, since a relatively rapid task may not give enough time for rapport to develop, we have designed a longitudinal component, whereby participants engage in slightly different but comparable tasks, spending 2 weeks on each of three tasks, for a total of 6 weeks, as we have done in prior research (Finkelstein, 2018;Madaio et al, 2018;Cassell, 2022). In order to collect high-fidelity data, the task needs to be displayed on the external screen in front of the child (rather than on a piece of cardboard laid on a table, for example) so that the child's face and body are oriented toward a webcam and microphone placed next to the monitor (the computer and keyboard will be located elsewhere to diminish distractions).…”

“…These children will engage only in solo work, and in a report to the teacher each week, with a pre-test and post-test as above, in order to differentiate between the impact on learning gains of simply engaging in the task and that of collaborating on that task. These tasks, annotation schemes and performance metrics, have been successfully used in both a child-child and child-VP longitudinal study with 7-to 8-year-olds (Finkelstein, 2018;Cassell, 2022), and have been successfully pilot-tested with 5-6 year-olds.…”

How About Kind of Generating Hedges using End-to-End Neural Models?

“…Since the 1990s, research has shown the the importance of intelligent tutoring systems as effective learning environment,s and supports for classroom learning (Anderson et al, 1995). Peer tutoring plays a powerful role as well, as peer tutors can motivate learners to try harder, as well as helping them to succeed, and it is particularly effective for low-achieving learners (Cassell, 2022). But virtual peer tutors have not yet achieved their potential, in part because of the difficulty of generating the social infrastructure of peer learning as well as the content of the matter being tutored.…”

Artificial Intelligence Enhances Children’s Science Learning from Television Shows

“…programming has now become possible with recent advances in conversational artificial intelligence including speech technologies and natural language processing. Young children already frequently talk to artificial intelligence (AI-) enabled devices or toys available in their homes-such as smart speakers, social robots, or Internet-connected stuffed animals (Cassell, 2022;Druga et al, 2018;Lovato et al, 2019)-and many of these applications were designed with specific learning or developmental goals in mind. The research has found that, while limited in its capacity in carrying out free-form, creative conversation with children, conversational AI is well suited for engaging in structured dialogue with children, which typically structed as asking children questions, listening in, and providing responsive feedback.…”