DynGL-SDP: Dynamic Graph Learning for Semantic Dependency Parsing

Proceedings of the 29th International Conference on Computational Linguistics, pages 3994–4004 October 12–17, 2022. 3994 DynGL-SDP: Dynamic Graph Learning for Semantic Dependency Parsing Bin Li1,2, Miao Gao 1,2, Yunlong Fan1,2, Yikemaiti Sataer1,2, Zhiqiang Gao1,2,∗, Yaocheng Gui3 1School of Computer Science and Engineering, Southeast University, Nanjing 210096, China 2Key Laboratory of Computer Network and Information Integration, Ministry of Education, Nanjing 210096, China 3School and Institute of Modern Posts, Nanjing University of Posts and Telecommunications, Nanjing 210003, China {lib, miaogao, fanyunlong, yikmat, zqgao}@seu.edu.cn guiyc@njupt.edu.cn Abstract A recent success in semantic dependency pars- ing shows that graph neural networks can make significant accuracy improvements, owing to its powerful ability in learning expressive graph representations. However, this work learns graph representations based on a static graph constructed by an existing parser, suffering from two drawbacks: (1) the static graph might be error-prone (e.g., noisy or incomplete), and (2) graph construction stage and graph represen- tation learning stage are disjoint, the errors in- troduced in the graph construction stage cannot be corrected and might be accumulated to later stages. To address these two drawbacks, we propose a dynamic graph learning framework and apply it to semantic dependency parsing, for jointly learning graph structure and graph representations. Experimental results show that our parser outperforms the previous parsers on the SemEval-2015 Task 18 dataset in three lan- guages (English, Chinese, and Czech). 1 Introduction Semantic dependency parsing (SDP) represents a sentence as a directed acyclic graph (DAG), also called semantic dependency graph (SDG), to cap- ture between-word semantic relationships that are more closely related to the meaning of the sentence. SDP has been widely applied in many downstream tasks of natural language processing (NLP), includ- ing sentiment analysis (Lin et al., 2019), abstractive summarization (Jin et al., 2020a), natural language understanding (Wu et al., 2021), etc. Several semantic dependency parsers are pre- sented in recent years. Their parsing mecha- nisms are either transition-based or graph-based. A transition-based parser generates a sequence of transition actions to incrementally build an SDG (Wang et al., 2018; Kurita and Søgaard, 2019; Lindemann et al., 2020; Fernández-González ∗Corresponding Author Dynamic Graph Graph Structure Learning buy Mary wants to Existing Parser book a Static Graph buy Mary wants to book a Mary wants to buy a book ARG2 BV root ARG1 ARG1 ARG1 Semantic Dependency Graph Mary wants to buy a book ARG2 BV root ARG1 ARG1 ARG1 Semantic Dependency Graph ARG3 GNN GNN Learning Learning (a) Static graph-based disjoint learning for SDP Dynamic Graph Graph Structure Learning buy Mary wants to Existing Parser book a Static Graph buy Mary wants to book a Mary wants to buy a book ARG2 BV root ARG1 ARG1 ARG1 Semantic Dependency Graph Mary wants to buy a book ARG2 BV root ARG1 ARG1 ARG1 Semantic Dependency Graph ARG3 GNN GNN Learning Learning (b) Dynamic graph-based joint learning for SDP Figure 1: Static graph-based SDP and dynamic graph- based SDP for example sentence "Mary wants to buy a book". The edge and dependency colored red are erroneous. …

Content truncated to 500 words