Identifying Discourse Boundaries in Group Discussions using Multimodal Features

Abstract

This study proposes models for detecting conversation boundaries in group discussions. First, we created a multimodal embedding space using an autoencoder, and applied a similarity-based approach to detect the discussion boundary. As the second method, we annotated conversation boundaries and created unimodal CNN models for language, audio, and head motion information. Then, created multimodal models by concatenating the output of unimodal models. In the evaluation experiment, we found that language information was the most useful modality, but by combining with audio and head motion modalities, the CNN-based models more accurately predict the conversation boundaries.