A Mental State Extraction Dataset for Theory-of-Mind-based Reasoning in Emotional Support Conversations

Abstract

Emotional Support Conversations (ESC) aim to both reduce users’ emotional distress and facilitate problem-solving. Recent approaches in ESC have explored incorporating commonsense knowledge into large language models (LLMs) to improve response generation. However, existing commonsense reasoning models often rely solely on the final utterance, fail to anticipate future turns, overlook emotional cues, or treat knowledge types independently, resulting in incoherent or emotionally misaligned responses. To address these limitations, we propose an approach grounded in Theory of Mind (ToM). Specifically, we introduce MENTOS, a dataset that provides turn-level annotations of the assistant’s mental states (Belief, Emotion, and Intent), organized in a causal structure reflecting psychological principles. A commonsense reasoning model trained on MENTOS predicts these mental states as intermediate reasoning signals that guide response generation. Experiments on the ESConv and ExTES datasets show that incorporating the inferred mental states can enhance supportive and goal-directed response generation across multiple reasoning backbones and response generators. Ablation studies further confirm that Belief, Emotion, and Intent provide complementary benefits for ESC tasks. These findings highlight the effectiveness of ToM-grounded intermediate reasoning in generating empathetic and contextually appropriate responses.

Resources

Details