E&S-Gainer: An Emotion Aware and Strategy Enhanced Model for Emotional Support Conversation

The Emotional Support Conversation (ESC) task aims to alleviate emotional distress in seekers and offer them an outlet for expressing their aggravation, garnering considerable attention in recent years. In ESC, accurately discerning the seeker’s emotional state is crucial, along with the strategic selection of appropriate support measures at various stages of the conversation to provide comfort. Previous methods fail to intricately delineate the objectives of the emotion support task, thereby hindering the supporter from crafting contextually fitting responses in each round, informed by both strategies and the user’s prevailing emotional state. To tackle this challenge, we propose a novel model, E&S-Gainer, leveraging LLaMA2’s inferential content from the conversation history to aid in training the model for improved emotional perception and strategy planning capabilities. Specifically, we devise few-shot and zero-shot prompts to harness LLaMA2 for generating stage-specific goals and pseudo-emotion labels, facilitating the separate training of the strategy-enhanced encoder and emotion-aware encoder. Additionally, we design a strategy-planning decoder explicitly incorporating strategies to guide response generation. Both automatic and human evaluations conducted on a prominent benchmark dataset attest to the superior performance of E&S-Gainer in terms of strategy selection and response generation. Our code is accessible at https://github.com/Whigle99/ES-Gainer .