Learning Time-Varying Turn-Taking Behavior in Group Conversations

TL;DR

The paper addresses predicting turn-taking in group conversations from individuals' traits and conversation history using a probabilistic framework that generalizes across groups. It introduces a time-varying proclivity model where the likelihood of speaking for each member is $P[s(t)=i] = u_i(t)/\mathbf{1}^\top \mathbf{u}(t)$ with $u_i(t) = [\pi_i + d_i w(\delta_i(t))] \mathbb{I}\{\delta_i(t)>1\}$ and learns $f$, $g$, and $\nu$ via neural nets to produce $\hat{u}_i(t)$. The methodology optimizes a log-likelihood objective across multiple groups, using block-coordinate descent and interpretable components to predict the next speaker, identify speakers, and generate conversations. Evaluations on synthetic and real data compare PRO to baselines (EXP, NM, HM), showing PRO better captures non-trivial turn-taking patterns, especially under the class-weighted $\ell_{\rm turn}$ loss. Real-world results reveal that actual group interactions exhibit slowly decaying or non-exponential proclivities, highlighting the need for flexible, data-driven yet principled models. Overall, the approach provides a generalizable and interpretable tool for analyzing and forecasting group communication dynamics with practical implications for team design and interactive systems.

Abstract

We propose a flexible probabilistic model for predicting turn-taking patterns in group conversations based solely on individual characteristics and past speaking behavior. Many models of conversation dynamics cannot yield insights that generalize beyond a single group. Moreover, past works often aim to characterize speaking behavior through a universal formulation that may not be suitable for all groups. We thus develop a generalization of prior conversation models that predicts speaking turns among individuals in any group based on their individual characteristics, that is, personality traits, and prior speaking behavior. Importantly, our approach provides the novel ability to learn how speaking inclination varies based on when individuals last spoke. We apply our model to synthetic and real-world conversation data to verify the proposed approach and characterize real group interactions. Our results demonstrate that previous behavioral models may not always be realistic, motivating our data-driven yet theoretically grounded approach.

Figures (2)

• Figure 1: (a) Testing loss for each baseline when $w = w_{{\rm exp}}$. The left five boxplots are measured via $\ell$ in \ref{['eq:loss']}, the right five via $\ell_{{\rm turn}}$ in \ref{['eq:ttloss']}. (b) Testing loss for each baseline when $w = w_{{\rm sig}}$. (c) Testing loss for each baseline using real-world conversation data.
• Figure 2: (a) True versus learned proclivities for either a sigmoidal proclivity $w_{{\rm sig}}$ or exponentially decaying proclivity $w_{{\rm exp}}$. (b) Learned proclivities for real-world conversation data.