Estimating performance of group interaction based on product dimension

Abstract

This paper focuses on developing a model for estimating the quality of discussion using multimodal features. For this purpose, we use a group meeting corpus including audio signal data of participants observed in 30 meeting sessions. Also, four annotators watch conversation transcripts and annotate the score about quality of discussion using product dimension which is a sociological criteria. We extracted various kinds of features such as spoken utterances, acoustic features, speaking turns. First, binary (high or low) classification models are trained to infer the annotated score from these features using support vector machine. Second, binary classification models are developed to infer the quality of unknown discussion task. Experiments results show that multimodal model archived 0.92 as the classification accuracy and task independent model archived 0.73.