Fast and Accurate End-to-End Span-based Semantic Role Labeling as Word-based Graph Parsing

Proceedings of the 29th International Conference on Computational Linguistics, pages 4160–4171 October 12–17, 2022. 4160 Fast and Accurate End-to-End Span-based Semantic Role Labeling as Word-based Graph Parsing Shilin Zhou, Qingrong Xia, Zhenghua Li∗, Yu Zhang, Yu Hong, Min Zhang Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, Suzhou, China {slzhou.cs,yzhang.cs}@outlook.com; kirosummer.nlp@gmail.com {zhli13,hongy,minzhang}@suda.edu.cn Abstract This paper proposes to cast end-to-end span- based SRL as a word-based graph parsing task. The major challenge is how to repre- sent spans at the word level. Borrowing ideas from research on Chinese word segmentation and named entity recognition, we propose and compare four different schemata of graph repre- sentation, i.e., BES, BE, BIES, and BII, among which we find that the BES schema performs the best. We further gain interesting insights through detailed analysis. Moreover, we pro- pose a simple constrained Viterbi procedure to ensure the legality of the output graph ac- cording to the constraints of the SRL struc- ture. We conduct experiments on two widely used benchmark datasets, i.e., CoNLL05 and CoNLL12. Results show that our word-based graph parsing approach achieves consistently better performance than previous results, under all settings of end-to-end and predicate-given, without and with pre-trained language mod- els (PLMs). More importantly, our model can parse 669/252 sentences per second, without and with PLMs respectively. 1 Introduction As a fundamental natural language processing (NLP) task, semantic role labeling (SRL) uses predicate-argument structure to represent the shal- low semantic meaning of sentences. SRL structure is shown to be helpful for many downstream NLP tasks, such as machine translation (Liu and Gildea, 2010; Marcheggiani et al., 2018) and question an- swering (Wang et al., 2015). There exist two forms of concrete SRL for- malism in the community, i.e., word-based (also known as dependency-based SRL) and span-based, depending on whether an argument consists of a single word or a word span. Compared with word-based SRL, span-based SRL is more com- plex due to difficulties in determining argument ∗ Corresponding author They want to do more . They want to do more . A0 A1 A0 A1 Figure 1: An example of span-based SRL, where “want” and “do” are two predicates. boundaries. Figure 1 shows the span-based SRL structure for two predicates. Semantic roles of ar- guments are distinguished with edge labels, such as “A0” (agent) and “A1” (patient). This work focuses on the end-to-end span-based SRL task, and pro- poses a unified model to simultaneously recognize predicates and arguments in the input sentence. In recent years, span-based SRL has achieved substantial performance boost due to the tremen- dous progress made by deep neural network mod- els, especially by pre-trained language models (PLMs). Currently, there are mainly two main- stream approaches, i.e., BIO-based (Zhou and Xu, 2015) sequence labeling and span-based graph pars- ing (He et al., 2018). The BIO-based sequence labeling approach first identifies the predicates and then finds arguments for each predicate independently by labeling every word with BIO tags, like “B-A0” or “I-A0”. Its major weakness is that …

Content truncated to 500 words