Graph Convolutions over Constituent Trees for Syntax-Aware Semantic Role Labeling

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 3915–3928, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 3915 Graph Convolutions over Constituent Trees for Syntax-Aware Semantic Role Labeling Diego Marcheggiani1∗ Ivan Titov2,3 1Amazon 2ILCC, School of Informatics, University of Edinburgh 3ILLC, University of Amsterdam marchegg@amazon.es ititov@inf.ed.ac.uk Abstract Semantic role labeling (SRL) is the task of identifying predicates and labeling argument spans with semantic roles. Even though most semantic-role formalisms are built upon con- stituent syntax, and only syntactic constituents can be labeled as arguments (e.g., FrameNet and PropBank), all the recent work on syntax- aware SRL relies on dependency represen- tations of syntax. In contrast, we show how graph convolutional networks (GCNs) can be used to encode constituent structures and inform an SRL system. Nodes in our SpanGCN correspond to constituents. The computation is done in 3 stages. First, ini- tial node representations are produced by ‘composing’ word representations of the first and last words in the constituent. Second, graph convolutions relying on the constituent tree are performed, yielding syntactically- informed constituent representations. Finally, the constituent representations are ‘decom- posed’ back into word representations, which are used as input to the SRL classifier. We eval- uate SpanGCN against alternatives, including a model using GCNs over dependency trees, and show its effectiveness on standard En- glish SRL benchmarks CoNLL-2005, CoNLL- 2012, and FrameNet. 1 Introduction The task of semantic role labeling (SRL) involves predicting the predicate-argument structure of a sentence. More formally, for every predicate, the SRL model must identify all argument spans and label them with their semantic roles (see Figure 1). The most popular resources for estimating SRL models are PropBank (Palmer et al., 2005) and FrameNet (Baker et al., 1998). In both cases, annotations are made on top of syntactic con- stituent structures. Earlier work on SRL hinged ∗Research conducted when the author was at the Univer- sity of Amsterdam. Investors appeal to the CEO not to limit their access to sales data V V ARG1 ARG0 AM-NEG ARG0 ARG1 ARG2 B-A0 V O B-A1 I-A1 B-A2 I-A2 I-A2 I-A2 I-A2 I-A2 I-A2 I-A2 O O O B-A0 I-A10 B-NEG O V B-A1 I-A1 I-A1 I-A1 I-A1 Figure 1: An example with semantic-role annotation and its reduction to the sequence labeling problem (BIO labels): the argument structure for predicates ap- peal and limit are shown in blue and red, respectively. on constituent syntactic structure, using the trees to derive features and constraints on role assign- ments (Gildea and Jurafsky, 2002; Pradhan et al., 2005; Punyakanok et al., 2008). In contrast, mod- ern SRL systems largely ignore treebank syntax (He et al., 2018a, 2017; Marcheggiani et al., 2017; Zhou and Xu, 2015) and instead use powerful fea- ture extractors, for example, LSTM sentence en- coders. There have been recent successful attempts to im- prove neural SRL models using syntax (Marcheg- giani and Titov, 2017; Strubell et al., 2018; He et al., 2018b). Nevertheless, they have relied on syntactic dependency representations rather than constituent trees. …

Content truncated to 500 words