EmoSum: Conversation Summarization with Emotional Consistency

Youngjin Jo, Jinyeong Bak

Abstract: Conversation summarization, a subcategory of text summarization that aims to extract key information from a conversation, is one of the most interesting research topics because it can help conversational artificial intelligence agents, such as chatbots, better understand the content of a conversation and thus improve the performance of conversational artificial intelligence agents. Conversation summarization requires consideration of various aspects of a conversation, including the topic of the conversation, who the speakers are, and what emotions they are feeling. However, despite these important aspects, most existing studies focus on providing only content-based summaries. This is a difficult problem to solve because learning all the important aspects of a conversation can lead to catastrophic forgetting. We introduce EmoSum, a conversation summarization model that can generate better summaries reflecting not only the content of the conversation but also the emotions of the conversation, which is one of the important aspects of the conversation. We also propose a simple but effective training method to solve the problem of catastrophic forgetting that may occur as the conversation summarization model learns more knowledge. Experimental results show that EmoSum trained by the proposed method outperforms baselines in generating more comprehensive and accurate summaries that reflect important aspects of the conversation, such as emotions and content.