Publications
WORKSHOP (DOMESTIC) ペルソナ一貫性の考慮と知識ベースを統合した HyperCLOVA を用いた雑談対話システム
山崎 天, 坂田 亘, 川本 稔己 (LINE/東京工業大学), 小林 滉河, Nguyen Tung, 上村 卓史, 中町礼文, 李聖哲, 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)
November 29, 2021
This paper describes a dialog system which we submitted to the Open Track of the Dialog System Live Competition 4. The system uses HyperCLOVA, a large-scale language model based on Transformer, and generates response candidates using Few-Shot prompts that are selected according to the user’s utterance. There are four types of prompts, each intended for general response generation, knowledge response generation, response generation considering persona consistency, and response generation for avoiding short sentences, which are used in combination. The system later filters the candidates to remove inappropriate utterances to obtain the final output. In the preliminary round, the system won the first place, showing that large-scale language models are effective in generating responses in open-domain dialogs. Meanwhile, several issues became apparent, for example the generation of offensive responses. In this paper, we describe the implementation and evaluation of our dialog system, and discuss the existing issues and future directions.
Paper :
ペルソナ一貫性の考慮と知識ベースを統合した HyperCLOVA を用いた雑談対話システム
(external link)