Soft engagement with pseudo initiatives for multi-party dialogue generation

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

1

초록

Multi-party conversations commonly take place on online discussion platforms as well as in various offline situations. For dialogue systems to effectively comprehend multi-party conversations, it is important to understand the dynamic nature of engagement between speakers and the unique characteristics of each speaker. However, recent approaches only emphasize the direct engagement of a speaker and addressee, leaving out other indirectly participating interlocutors. In this work, we present a simple yet effective method, SoPI, which introduces two novel modules to capture the subtle variations in engagement level. We also introduce the concept of pseudo initiatives that are extracted from the dialogue context as a key element in multi-party dialogue generation. Extensive experiments are conducted on the UbuntuIRC benchmark and SoPI achieves the best performance in all metrics including human evaluation, compared to the previous state-of-the-art model and various baselines. Analysis on the effect of the varying number of interlocutors further show the effectiveness of SoPI in Multi-party conversations.

키워드

Multi-party conversationResponse generationFocus learningSegment embedding
제목
Soft engagement with pseudo initiatives for multi-party dialogue generation
저자
Chang, YoungjaeKo, Youngjoong
DOI
10.1016/j.patrec.2025.03.001
발행일
2025-05
유형
Article
저널명
Pattern Recognition Letters
191
페이지
103 ~ 109