Prompt Design Using Past Dialogue Summarization for LLMs to Generate the Current Appropriate Dialogue

Recent technological innovations in large language models (LLMs) produce incredible performance. This also has a similar impact on dialogue systems. However, following fluently current dialogue from the past dialogue is crucial, especially for chat-oriented dialogue systems, which are difficult for only LLMs to handle. In this paper, we propose a prompt design using a method summarizing dialogue for LLMs to generate the current appropriate dialogue in chat-oriented dialogue systems. For dialogue summarization, we first use a hand-crafted dialogue summarization corpus and two other corpora, and then a language model that summarizes dialogue in several sentences is fine-tuned on the combined corpora. We conducted two experiments for the performance evaluation of the proposed method. One is to evaluate how much the constructed model summarizes dialogue in some patterns. Another is to evaluate a performance predicting the current dialogue by prompting an LLM using the summarization model in contrast to the whole past dialogue. Through all the evaluation, the results suggest that the proposed prompt design is useful for dialogue generation using LLMs.