ToMESC: A Theory of Mind-based Dataset for Emotional Support Conversations

Task-oriented dialogue systems aim to help users achieve specific goals through natural language interaction. Emotional Support Conversations (ESC), a representative task in this domain, focus on reducing users’ emotional distress while facilitating problem-solving. Recent approaches add commonsense knowledge to prompts for large language models (LLMs) to generate supportive responses. However, they often over-rely on the final utterance, neglect dialogue dynamics, fail to track emotional cues, or treat knowledge types independently—resulting in incoherent or emotionally misaligned outputs. To address these issues, we introduce ToMESC, a Theory of Mind-based dataset that models the latent mental states—Belief, Emotion, Desire, and Intent—of an emotionally supportive assistant in a structured causal sequence. Built upon the ESConv dataset, ToMESC provides turn-level annotations that map mental states to the evolving dialogue context. This structure enables the generation of more coherent and strategically aligned responses. We use ToMESC to train a model that infers the assistant's mental states from dialogue. These predicted states are then used as intermediate reasoning steps to guide response generation in a zero-shot LLM framework. Experiments demonstrate that fine-tuning on ToMESC enhances mental state inference, yielding responses that are more emotionally appropriate and goal-directed than those generated by existing commonsense reasoning baselines.面向任务的对话系统旨在通过自然语言交互帮助用户达成特定目标。情感支持对话（Emotional Support Conversations, ESC）作为该领域的代表性任务，核心在于在助力问题解决的同时缓解用户的情绪困扰。近期研究通过为大语言模型（Large Language Model, LLM）的提示词添加常识知识，以生成具有支持性的回复。然而此类方法往往过度依赖对话的最终话语，忽视对话动态演化过程，无法追踪情绪线索，或是对各类知识类型进行独立处理，最终生成的回复往往缺乏连贯性且与情绪适配要求相悖。为解决上述问题，我们提出了ToMESC：一款基于心理理论（Theory of Mind）的数据集，该数据集以结构化因果序列的形式，对情感支持助手的潜在心理状态——信念（Belief）、情绪（Emotion）、欲望（Desire）与意图（Intent）——进行建模。该数据集基于ESConv数据集构建，提供了轮级标注，将心理状态与不断演化的对话上下文进行关联映射。这一结构化设计有助于生成更具连贯性且策略适配的回复。我们利用ToMESC训练了一个模型，该模型可从对话中推断出助手的心理状态。随后，这些预测得到的心理状态将作为中间推理步骤，在零样本（Zero-shot）大语言模型框架中指导回复生成。实验结果表明，在ToMESC上进行微调能够提升心理状态推断的效果，所生成的回复相较于现有常识推理基线模型的输出，在情绪适配性与目标导向性上均更为出色。