Do you mind if I ask? Addressing the cold start problem in personalised relational agent conversation

To personalise dialogue to different users, relational agents need to learn about the users' preferences for relational cues used by the agent. In the context of a virtual advisor to reduce students' study stress, we designed a between-subjects study with three groups (empathic, neutral and adaptive) who either received all cues, no cues, or helpful cues only, respectively, and compared rapport and changes in study stress scores. To avoid the cold start problem, we sought to train the agent and adapt its dialogue to include or exclude 10 relational cues based on the user's responses to whether an example of each relational cue is found helpful prior to the session with the virtual advisor. The results of an experiment with 111 students show that the rapport scores for the empathic and adaptive groups were significantly higher than the neutral group; change in rapport scores was significantly higher in the adaptive group than in the empathic group. Furthermore, study stress scores significantly reduced for the adaptive and empathic groups, but not for the neutral group. We found some relationships between the number of times students found helpful what they received and other variables. We also found that the number of discrepancies and matches between what relational cues users received and what they found helpful were greatest in the adaptive group. This indicates the effectiveness of this approach for dealing with the cold start problem.