GIFT: Graph-Induced Fine-Tuning for Multi-Party Conversation Understanding

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11645–11658 July 9-14, 2023 ©2023 Association for Computational Linguistics GIFT: Graph-Induced Fine-Tuning for Multi-Party Conversation Understanding Jia-Chen Gu1, Zhen-Hua Ling1∗, Quan Liu2,3, Cong Liu1,3, Guoping Hu2,3 1National Engineering Research Center of Speech and Language Information Processing, University of Science and Technology of China, Hefei, China 2State Key Laboratory of Cognitive Intelligence 3iFLYTEK Research, Hefei, China {gujc,zhling}@ustc.edu.cn, {quanliu,congliu2,gphu}@iflytek.com Abstract Addressing the issues of who saying what to whom in multi-party conversations (MPCs) has recently attracted a lot of research attention. However, existing methods on MPC under- standing typically embed interlocutors and utterances into sequential information flows, or utilize only the superficial of inherent graph structures in MPCs. To this end, we present a plug-and-play and lightweight method named graph-induced fine-tuning (GIFT) which can adapt various Transformer-based pre-trained language models (PLMs) for universal MPC understanding. In detail, the full and equiva- lent connections among utterances in regular Transformer ignore the sparse but distinctive dependency of an utterance on another in MPCs. To distinguish different relationships between utterances, four types of edges are designed to integrate graph-induced signals into attention mechanisms to refine PLMs originally designed for processing sequential texts. We evaluate GIFT by implementing it into three PLMs, and test the performance on three downstream tasks including addressee recognition, speaker identification and response selection. Experimental results show that GIFT can significantly improve the performance of three PLMs on three downstream tasks and two benchmarks with only 4 additional parameters per encoding layer, achieving new state-of-the- art performance on MPC understanding. 1 Introduction Maintaining appropriate human-computer conver- sation is an important task leaping towards ad- vanced artificial intelligence. Most of existing methods have studied understanding conversations between two participants, aiming at returning an appropriate response either in a generation-based (Shang et al., 2015; Serban et al., 2016; Zhang et al., 2020; Roller et al., 2021) or retrieval-based manner (Wu et al., 2017; Zhou et al., 2018; Tao et al., 2019; ∗Corresponding author. : reply-to : replied-by : reply-self : indirect-reply U1 U2 U3 U4 U5 U6 U7 U8 U1 U2 U3 U4 U5 U6 U7 U8 (a) A Graphical Information Flow of an MPC (b) Reply Relationships in a Graph Structure for U3 : reply-to Figure 1: Illustration of (a) a graphical information flow of an MPC where rectangles denote utterances, and solid lines represent the “reply" relationship between two utterances, and (b) the detailed reply relationships between each utterance and U3. Gu et al., 2020). Recently, researchers have paid more attention to a more practical and challenging scenario involving more than two participants, which is well known as multi-party conversations (MPCs) (Ouchi and Tsuboi, 2016; Zhang et al., 2018; Le et al., 2019; Hu et al., 2019; Wang et al., 2020; Gu et al., 2021, 2022). Unlike two- party conversations, utterances in an MPC can be spoken by anyone and address anyone else in this conversation, constituting a graphical information flow and various relationships between …

Content truncated to 500 words