DiaRoBERTa: A Multi-Party Dialogue Model for Multi-Skill Recognition in Classroom Collaborative Problem Solving Discussions

Abstract

Collaborative problem-solving (CPS) integrates cognitive and social dimensions, which makes it a critical skill in educational contexts. However, existing models often treat dialogues as linear sequences, which limits their ability to capture the graph-like structures and intricate interrelations inherent in CPS discussions. These limitations hinder the accurate representation of speaker roles, conversational transitions, and dependencies within multi-party interactions. To address this challenge, we propose DiaRoBERTa, a novel model tailored for CPS dialogue classification. Building on Multi-Party Conversation (MPC) techniques, DiaRoBERTa leverages the ability of MPC to model dialogues as interconnected graphs rather than as mere linear sequences. Specifically, MPC techniques allow us to incorporate specialized markers that encode speaker roles and conversational transitions, thereby explicitly capturing the complex relational dynamics among participants. This enables the model not only to maintain the sequential order of utterances but also to represent the non-linear dependencies and interaction patterns critical for understanding CPS dialogues. Additionally, a targeted loss function addresses class imbalance, improving performance on underrepresented categories.Experiments on a dataset of middle school classroom discussions demonstrate that DiaRoBERTa outperforms baseline models, achieving a 1.5% improvement over existing methods. These results highlight DiaRoBERTa ’ s effectiveness in adapting sequence-based approaches to handle graph-like structures and its scalability for CPS skill assessment. By addressing key limitations of current methods, DiaRoBERTa provides a robust framework for advancing automated CPS dialogue analysis in collaborative educational settings.