Predicting Dialogue Breakdown in Conversational Pedagogical Agents with Multimodal LSTMs

摘要

Recent years have seen a growing interest in conversational pedagogical agents. However, creating robust dialogue managers for conversational pedagogical agents poses significant challenges. Agents' misunderstandings and inappropriate responses may cause breakdowns in conversational flow, lead to breaches of trust in agent-student relationships, and negatively impact student learning. Dialogue breakdown detection (DBD) is the task of predicting whether an agent's utterance will cause a breakdown in an ongoing conversation. A robust DBD framework can support enhanced user experiences by choosing more appropriate responses, while also offering a method to conduct error analyses and improve dialogue managers. This paper presents a multimodal deep learning-based DBD framework to predict breakdowns in student-agent conversations. We investigate this framework with dialogues between middle school students and a conversational pedagogical agent in a game-based learning environment. Results from a study with 92 middle school students demonstrate that multimodal long short-term memory network (LSTM)-based dialogue breakdown detectors incorporating eye gaze features achieve high predictive accuracies and recall rates, suggesting that multimodal detectors can play an important role in designing conversational pedagogical agents that effectively engage students in dialogue.