Graph-Based Prediction of Meeting Participation

Gabriel Murray

doi:10.3390/mti3030054

Abstract

Given a meeting participant’s turn-taking dynamics during one segment of a meeting, and their contribution to the group discussion up to that point, our aim is to automatically predict their activity level at a later point of the meeting. The predictive models use verbal and nonverbal features derived from social network representations of each small group interaction. The best automatic prediction models consistently outperform two baseline models at multiple time-lags. We analyze which interaction features are most predictive of later meeting activity levels, and investigate the efficacy of the verbal vs. nonverbal feature classes for this prediction task. At long time-lags, linguistic features become more crucial, but performance degrades compared with prediction at short time-lags.

Highlights

The core question in this research is whether we can automatically predict a conversation participant’s activity levels at a later point in the conversation, given features that describe their current turn-taking dynamics and contributions to the discussion so far
We have addressed the task of automatically predicting a group member’s participation level at a later point in a meeting, given features representing their current turn-taking dynamics and their contribution to the conversation so far
Most of the features are derived from graph representations of turn-taking and of the language used in the conversation

Summary

Introduction

The core question in this research is whether we can automatically predict a conversation participant’s activity levels at a later point in the conversation, given features that describe their current turn-taking dynamics and contributions to the discussion so far. Such predictions could be useful for an automated meeting assistant that provides feedback on interaction patterns, with the goal of encouraging participation. The machine learning models used for this task utilize both verbal and nonverbal features of the discussion Both classes of features are derived from graph-based representations of the conversation, and the features include measures of network centrality and network change. The verbal features capture aspects of topic structure, linguistic alignment, and lexical change during the conversation, based on a tripartite graph representation linking speakers, words, and sentences

Objectives

Methods

Results

Conclusion