Autoregressive cross-interlocutor attention scores meaningfully capture conversational dynamics

This paper analyzes attention scores over a conversational partner's historical turns trained with an autoregressive prosodic objective. Following a qualitative observation that these attention scores seem to organize dialogue history into topic segments, we demonstrate that they capture meaningful dialogue structure based on several quantitative measures. This finding has implications for spoken dialogue system design and analysis of entrainment and conversational dynamics in human-human and human-machine communication.