Emotionally intelligent conversational agents: from understanding to interaction

Abstract

To advance artificial intelligence, it is essential to equip machines with emotional intelligence, thereby enhancing human-AI communication and relationships. In this thesis, I present our work on building emotionally intelligent conversational agents, focusing on three key themes: empathetic understanding, reliable responding, and engaging interaction. For empathetic understanding, we propose two models, GREC and CARE, which are designed to generate empathetic responses by interpreting user emotions and the underlying emotional causalities through graph-based structures. While GREC reasons over an external commonsense knowledge graph, CARE integrates causal relationship inference directly within the model. To ensure reliable responding, we address two challenges. First, we introduce d-PM, a method to learn user preferences while accounting for individual disagreements, and align conversational agents accordingly. Second, to mitigate unhelpful responses that could hinder emotional support, we propose Muffin, a framework that reduces the likelihood of such responses by leveraging multi-faceted AI feedback. These two works are complementary, where one increases user satisfaction and the other mitigates unhelpfulness. Both methods are model-agnostic and can enhance transformer-based models, including state-of-the-art ones. The last theme centers on engaging interaction in emotionally intelligent conversational agents. We present two works: one for evaluation and one for model alignment. Since conversation engagement reflects the overall experience of an entire dialogue and involving real human users can be costly, we employ model-simulated users in our studies. First, we propose ClientCAST, a framework to evaluate LLM-based therapists. After interacting with the conversational agents, simulated clients complete questionnaires to assess the overall conversational engagement. Second, to enhance engagement, we align conversational agents with conversations that are likely to produce higher engagement levels. This is achieved through Monte Carlo Tree Search for interaction, which identifies dialogue trajectories associated with greater user engagement. Together, these contributions offer a comprehensive approach to building emotionally intelligent conversational agents that are empathetic, reliable, and engaging.