Enhancing dialogue response generation through emotion memory

Abstract

Open-Domain Dialogue (ODD) systems present unique challenges in generating engaging conversations such as contextual understanding and generating engaging and natural responses. With the recent remarkable progress of Large Language Models (LLM), we aim to enhance smaller dialogue models by extracting emotion knowledge from LLMs. We introduce the concept of Emotion-aware Memory (EM), which acts as a repository to store key emotional features extracted from LLMs. By leveraging EM during the response decoding stage, we aim to generate emotionally-enriched responses. We experiment and achieve state-of-the-art performance on DailyDialog and EmpatheticDialogues datasets. This research contributes to emotionally-aware dialogue systems and paves the way for more engaging conversational agents.