Enhancing Zero-shot Emotion Perception in Conversation through the Internal-to-External Chain-of-Thought
Abstract: An excellent emotional dialogue model needs to rapidly adapt to new scenarios and perform emotion analysis to meet rapidly changing demands. Therefore, enhancing the model's zero-shot emotion-related capabilities in the dialogue domain has become a new challenge. However, current research shows that large language models (LLMs) perform poorly in zero-shot emotion-related tasks and the Emotion Recognition in Conversations (ERC) task alone doesn't comprehensively reflect the model's emotion understanding capabilities. In this paper, we propose an Emotion Perception in Conversation (EPC) task, which includes both ERC and Emotion Inference in Conversations (EIC), to evaluate the model's emotion perception capabilities in dialogue comprehensively. We propose an Internal-to-External Chain-of-Thought (IoECoT) method for the EPC task. This is a plug-and-play method that first extracts personality information of the dialogue participants from the dialogue history as internal factors influencing emotions, and then uses the sentiment polarity of the historical utterances as external factors. Finally, emotions are perceived by combining internal and external factors. Additionally, we conduct extensive experiments, and the results show that IoECoT significantly outperforms other baselines across multiple models and datasets, demonstrating that IoECoT effectively enhances the emotion perception capabilities of LLMs in zero-shot scenarios.
Paper Type: Long
Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining
Research Area Keywords: emotion recognition in conversations; emotion inference in conversations; chain-of-thought; zero-shot
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 5372
Loading