A Simple and Accurate Syntax-Agnostic Neural Model for Dependency-based Semantic Role Labeling

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 411–420, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics A Simple and Accurate Syntax-Agnostic Neural Model for Dependency-based Semantic Role Labeling Diego Marcheggiani1, Anton Frolov2, Ivan Titov1,3 1ILLC, University of Amsterdam 2Machine Intelligence Department, Yandex 3ILCC, School of Informatics, University of Edinburgh marcheggiani@uva.nl anton-fr@yandex-team.ru ititov@inf.ed.ac.uk Abstract We introduce a simple and accurate neu- ral model for dependency-based seman- tic role labeling. Our model predicts predicate-argument dependencies relying on states of a bidirectional LSTM en- coder. The semantic role labeler achieves competitive performance on English, even without any kind of syntactic information and only using local inference. How- ever, when automatically predicted part- of-speech tags are provided as input, it substantially outperforms all previous lo- cal models and approaches the best re- ported results on the English CoNLL- 2009 dataset. We also consider Chi- nese, Czech and Spanish where our approach also achieves competitive re- sults. Syntactic parsers are unreliable on out-of-domain data, so standard (i.e., syntactically-informed) SRL models are hindered when tested in this setting. Our syntax-agnostic model appears more ro- bust, resulting in the best reported results on standard out-of-domain test sets. 1 Introduction The task of semantic role labeling (SRL), pio- neered by Gildea and Jurafsky (2002), involves the prediction of predicate argument structure, i.e., both identification of arguments as well as their as- signment to an underlying semantic role. These representations have been shown to be benefi- cial in many NLP applications, including ques- tion answering (Shen and Lapata, 2007) and in- formation extraction (Christensen et al., 2011). Semantic banks (e.g., PropBank (Palmer et al., 2005)) often represent arguments as syntactic con- stituents or, more generally, text spans (Baker Sequa makes and repairs jet engines. 01 01 01 A0 A1 A0 A1 A1 Figure 1: A semantic dependency graph. et al., 1998). In contrast, CoNLL-2008 and 2009 shared tasks (Surdeanu et al., 2008; Hajic et al., 2009) popularized dependency-based se- mantic role labeling where the goal is to iden- tify syntactic heads of arguments rather than entire constituents. Figure 1 shows an example of such a dependency-based representation: node labels are senses of predicates (e.g., “01” indicates that the first sense from the PropBank sense repository is used for predicate makes in this sentence) and edge labels are semantic roles (e.g., A0 is a proto-agent, ‘doer’). Until recently, state-of-the-art SRL systems relied on complex sets of lexico-syntactic fea- tures (Pradhan et al., 2005) as well as declara- tive constraints (Punyakanok et al., 2008; Roth and Yih, 2005). Neural SRL models instead ex- ploited feature induction capabilities of neural net- works, largely eliminating the need for complex hand-crafted features. Initially achieving state- of-the-art results only in the multilingual setting, where careful feature engineering is not practi- cal (Gesmundo et al., 2009; Titov et al., 2009), neural SRL models now also outperform their tra- ditional counterparts on standard benchmarks for English (FitzGerald et al., 2015; Roth and Lapata, 2016; Swayamdipta et …

Content truncated to 500 words