[4E3-GS-8] Robot and real worlds:

[4E3-GS-8-04] Real-time Open-domain Dialogue System using an Android Robot Based on Two Types of Dialogue Summarization for a Large Language Model

〇Ryota Fukui¹, Akishige Yuguchi¹, Yoshio Matsumoto¹, Yuya Okadome¹ (1. Tokyo University of Science)

Keywords:android robot, open-domain dialogue, large language models, prompt design

Recent technological innovations in large language models (LLMs) enable conversational robots to generate human-like responses. In this paper, we propose a real-time LLM-based open-domain dialogue system using an android robot to talk with a human for a long time. The dialogue system adopts a prompt design based on two types of dialogue summarization for LLMs. The first type is the abstractive dialogue summarization using the fine-tuned language model with several summarization corpora. The second type is the noun phrase extraction from a user’s past utterances. This prompt design can be expected to compress the past dialogue contexts for the LLM-based conversational system to generate the appropriate current utterance to maintain a long-term opendomain conversation. We compare the proposed conversational android system to the system without the past dialogue contexts in the subjective evaluation. The results showed that the subjects felt a significant impression that the system understood their talks.