Publications

ワークショップ (国内) HyperCLOVA を利用したプロンプトプログラミングによるシチュエーションに適した応答生成

川本 稔己 (LINE/東京工業大学), 山崎 天, 坂田 亘, 佐藤 敏紀

第12回 対話システムシンポジウム (第12回 対話システムシンポジウム)

2021.11.29

This paper describes our dialogue system submitted to the Situation Track of the Dialogue System Live Competition 4. Our system generates responses by using HyperCLOVA, a language model based on Transformer. Prompts are given not only the dialogue history but also the description of situation and personas to enable situation-based response generation. Moreover, our system has a mechanism to regenerate a response with adding starting words of the response in case the response does not match the situation. As a result, our system placed second in the preliminary round. We confirmed that the large-scale Japanese language model is useful for the task of situation-based dialogue generation. On the other hand, the reliability and the consistency of dialogues remains an issue because the hallucination problem often occurred. In this paper, we discuss the issues by referring to dialogue logs.

Paper : HyperCLOVA を利用したプロンプトプログラミングによるシチュエーションに適した応答生成新しいタブまたはウィンドウで開く (外部サイト)