How to Best Use Syntax in Semantic Role Labelling

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5338–5343 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5338 How to best use Syntax in Semantic Role Labelling Yufei Wang1 and Mark Johnson1 and Stephen Wan2 and Yifang Sun3 and Wei Wang3 Macquarie University, Sydney, Australia1 CSIRO Data61, Sydney, Australia2 The University of New South Wales, Sydney, Australia3 yufei.wang@students.mq.edu.au, mark.johnson@mq.edu.au stephen.wan@data61.csiro.au {yifangs,weiw}@cse.unsw.edu.au Abstract There are many different ways in which exter- nal information might be used in an NLP task. This paper investigates how external syntac- tic information can be used most effectively in the Semantic Role Labeling (SRL) task. We evaluate three different ways of encod- ing syntactic parses and three different ways of injecting them into a state-of-the-art neural ELMo-based SRL sequence labelling model. We show that using a constituency represen- tation as input features improves performance the most, achieving a new state-of-the-art for non-ensemble SRL models on the in-domain CoNLL’05 and CoNLL’12 benchmarks.1 1 Introduction Properly integrating external information into neu- ral networks has received increasing attention re- cently (Wu et al., 2018; Li et al., 2017; Strubell et al., 2018). Previous research on this topic can be roughly categorized into three classes: i) In- put: The external information are presented as ad- ditional input features (i.e., dense real-valued vec- tors) to the neural network (Collobert et al., 2011). ii) Output: The neural network is trained to pre- dict the main task and the external information in a multi-task approach (Changpinyo et al., 2018). iii) Auto-encoder: This approach, recently proposed by Wu et al. (2018), simultaneously combines the Input and Output during neural models training. The simplicity of these methods allow them to ap- ply to many NLP sequence tasks and various neu- ral model architectures. However, previous studies often focus on inte- grating word-level shallow features such as POS or chunk tags into the sequence labelling tasks. Syntactic information, which encodes the long- range dependencies and global sentence structure, has not been studied as carefully. This paper fills 1Our model source code is available in https:// github.com/GaryYufei/bestParseSRL this gap by integrating syntactic information to the sequence labelling task. We address three ques- tions: 1) How should syntactic information be en- coded as word-level features? 2) What is the best way of integrating syntactic information? and 3) What effect does the choice of syntactic represen- tation have on the performance? We study these questions in the context of Se- mantic Role Labelling (SRL). A SRL system ex- tracts the predicate-argument structure of a sen- tence.2 Syntax was an essential component of early SRL systems (Xue and Palmer, 2004; Pun- yakanok et al., 2008). The state-of-the-art neu- ral SRL systems use a neural sequence labelling model without any syntax knowledge (He et al., 2018, 2017; Tan et al., 2018). We show below that injecting external syntactic knowledge into a neu- ral SRL sequence labelling model can improve the performance, and our best model sets a new state- of-the-art for …

Content truncated to 500 words