Existing studies on emotion analysis in conversations have mainly focused on recognizing the emotion of a given utterance. This paper investigates the task of emotion inference in conversations, which explores how the utterances affect the addressee’s emotion, without knowing the addressee’s response yet. While it is straightforward for humans to perceive and reason about the feelings of others in conversations, it is a severe challenge for machines, mainly due to the lack of commonsense knowledge. In this work, we propose to leverage external inferential knowledge to enhance the emotion inference in conversations. Specifically, a conversation modeling module is designed to accumulate information from the conversation history based on the emotional interaction between the addressee and writers. In addition, a knowledge integration strategy is also proposed to integrate the conversation-related commonsense knowledge generated from the event-based knowledge graph. The experiments on three different benchmark conversational datasets demonstrate the effectiveness of the proposed models, and prove the benefits of commonsense knowledge for emotion inference in conversations.

现有对对话中情感分析的研究主要集中在识别给定话语的情感上。本文研究了对话中情绪推理的任务，探索了话语如何影响受话者的情绪，而不知道受话者的反应。虽然人类在对话中感知和推理他人的感受很简单，但对机器来说却是一个严峻的挑战，主要是由于缺乏常识知识。在这项工作中，我们建议利用外部推理知识来增强对话中的情感推理。具体而言，对话建模模块旨在根据收件人和作者之间的情感互动从对话历史中积累信息。此外，还提出了一种知识集成策略，以集成从基于事件的知识图生成的对话相关常识知识。在三个不同的基准会话数据集上的实验证明了所提出模型的有效性，并证明了常识知识对会话中情感推理的好处。