Using Group History to Identify Character-Directed Utterances in Multi-Child Interactions.

Addressee identification is an element of all language-based interactions, and is critical for turn-taking. We examine the particular problem of identifying when each child playing an interactive game in a small group is speaking to an animated character. After analyzing child and adult behavior, we explore a family of machine learning models to integrate audio and visual features with temporal group interactions and limited, task-independent language. The best model performs identification about 20% better than the model that uses the audio-visual features of the child alone.