Syntax-aware Neural Semantic Role Labeling with Supertags

Proceedings of NAACL-HLT 2019, pages 701–709 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 701 Syntax-aware Neural Semantic Role Labeling with Supertags Jungo Kasai♣∗ Dan Friedman♠∗ Robert Frank♦ Dragomir Radev♦ Owen Rambow♥ ♣University of Washington ♠Google ♦Yale University ♥Elemental Cognition, LLP jkasai@cs.washington.edu danfriedman@google.com {robert.frank,dragomir.radev}@yale.edu owenr@elementalcognition.com Abstract We introduce a new syntax-aware model for dependency-based semantic role labeling that outperforms syntax-agnostic models for En- glish and Spanish. We use a BiLSTM to tag the text with supertags extracted from depen- dency parses, and we feed these supertags, along with words and parts of speech, into a deep highway BiLSTM for semantic role la- beling. Our model combines the strengths of earlier models that performed SRL on the ba- sis of a full dependency parse with more re- cent models that use no syntactic information at all. Our local and non-ensemble model achieves state-of-the-art performance on the CoNLL 09 English and Spanish datasets. SRL models benefit from syntactic information, and we show that supertagging is a simple, power- ful, and robust way to incorporate syntax into a neural SRL system. 1 Introduction Semantic role labeling (SRL) is the task of iden- tifying the semantic relationships between each predicate in a sentence and its arguments (Gildea and Jurafsky, 2002). While early research as- sumed that SRL models required syntactic infor- mation to perform well (Punyakanok et al., 2008), recent work has demonstrated that neural networks can achieve competitive and even state-of-the-art performance without any syntactic information at all (Zhou and Xu, 2015; Marcheggiani et al., 2017; He et al., 2017). These systems have the bene- fits of being simpler to implement and perform- ing more robustly on foreign languages and out- of-domain data, cases where syntactic parsing is more difficult (Marcheggiani et al., 2017). In this paper, we show that using supertags is an effective middle ground between using full syn- tactic parses and using no syntactic information ∗Work partially done at Yale University. at all. A supertag is a linguistically rich descrip- tion assigned to a lexical item. Supertags impose complex constraints on their local context, so su- pertagging can be thought of as “almost parsing” (Bangalore and Joshi, 1999). Supertagging has been shown to facilitate Tree-Adjoining Grammar (TAG) parsing (Bangalore et al., 2009; Friedman et al., 2017; Kasai et al., 2017, 2018) and Combi- natory Categorial Grammar (CCG) parsing (Clark and Curran, 2007; Kummerfeld et al., 2010; Lewis et al., 2016; Xu, 2016). We propose that supertags can serve as a rich source of syntactic information for downstream tasks without the need for full syntactic parsing. Following Ouchi et al. (2014), who used supertags to improve dependency parsing, we extract var- ious forms of supertags from the dependency- annotated CoNNL 09 corpus. This contrasts with prior SRL work that uses TAG or CCG supertags (Chen and Rambow, 2003; Lewis et al., 2015). We train a bidirectional LSTM (BiLSTM) to pre- dict supertags and feed the predicted supertag em- bedding, along with word and predicted part-of- speech …

Content truncated to 500 words