CSAGN: Conversational Structure Aware Graph Network for Conversational Semantic Role Labeling

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2312–2317 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2312 CSAGN: Conversational Structure Aware Graph Network for Conversational Semantic Role Labeling Han Wu1,2, Kun Xu3, Linqi Song1,2∗ 1City University of Hong Kong Shenzhen Research Institute 2Department of Computer Science, City University of Hong Kong 3Tencent AI Lab hanwu32-c@my.cityu.edu.hk kxkunxu@tencent.com linqi.song@cityu.edu.hk Abstract Conversational semantic role labeling (CSRL) is believed to be a crucial step towards dia- logue understanding. However, it remains a major challenge for existing CSRL parser to handle conversational structural information. In this paper, we present a simple and effec- tive architecture for CSRL which aims to ad- dress this problem. Our model is based on a conversational structure-aware graph network which explicitly encodes the speaker depen- dent information. We also propose a multi-task learning method to further improve the model. Experimental results on benchmark datasets show that our model with our proposed train- ing objectives significantly outperforms previ- ous baselines. 1 Introduction Recent research has achieved impressive improve- ments on conversation-based tasks, such as dia- logue response generation (Li et al., 2017; Dinan et al., 2019; Wu et al., 2019), task-oriented dia- logue modeling (Mrkši´c et al., 2017; Budzianowski et al., 2018) and conversational question answering (Choi et al., 2018; Reddy et al., 2019). However, the frequent occurrences of ellipsis and anaphora in human conversations still create huge challenges for dialogue understanding. To address this, Xu et al. (2021) proposed the Conversational Seman- tic Role Labeling (CSRL) task whose goal is to extract predicate-argument structures across the en- tire conversation. Figure 1 illustrates an example, where a CSRL parser needs to identify “《泰坦尼克 号》(Titanic)” as the ARG1 argument of the predi- cate “看过 (watched)" and the ARG0 argument of the predicate “是 (is)". One can see that in the origi- nal conversation, “《泰坦尼克号》(Titanic)” is omit- ted in the second turn and referred as “这 (this)" in the last turn. Xu et al. (2021) has demonstrated ∗Corresponding author. A: � (Have ) �� (watched) �� (movie) ARG0 ARG1 ARG0 ARG1 ����� (Titanic) � ? B: (� (Have not) �� (watched) , �������� �What kind of the movie is it� ? A: � (This) � (is) ���)��� (an American disaster film) � Figure 1: A conversational SRL example. the usefulness of CSRL on many downstream tasks such as dialogue generation and dialogue rewriting. Despite the successes that CSRL has achieved, existing CSRL model (Xu et al., 2021) is a simple extension of BERT. Specifically, they first encode each utterance into local contextual representations with pre-trained language models, and then utilize a stack of self-attention layers to obtain global con- textual representations. We argue that their model may suffer two main problems. First, in the local feature extraction phase, they ignore the fact that jointly considering the predicate and context utter- ances could help the model better identify some relevant ommited arguments. Second, in the global feature extraction phase, some vital conversational structural information, such as the speaker …

Content truncated to 500 words