Automatic Role Estimation of Students’ Utterance Using Generative AI

Abstract

This study explores the use of generative AI to estimate speaking roles in educational group discussions. By analyzing Zoom-recorded conversations transcribed into text, a modified ChatGPT model classified utterances into roles such as questions or assertions. Initial accuracy (50–55%) improved to 76% through role-specific training and contextual embedding. This method reduces manual effort and evaluator bias while enhancing scalability and objectivity in assessing collaborative learning. Tested across languages and cultures, it shows promise for international application. The approach also advances discourse analysis capabilities, contributing to learning analytics and supporting interventions to improve educational outcomes.