Joint A* CCG Parsing and Semantic Role Labelling

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1444–1454, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Joint A∗ CCG Parsing and Semantic Role Labeling Mike Lewis Luheng He Luke Zettlemoyer Computer Science & Engineering University of Washington Seattle, WA 98195 {mlewis,luheng,lsz}@cs.washington.edu Abstract Joint models of syntactic and semantic parsing have the potential to improve performance on both tasks—but to date, the best results have been achieved with pipelines. We introduce a joint model us- ing CCG, which is motivated by the close link between CCG syntax and semantics. Semantic roles are recovered by labelling the deep dependency structures produced by the grammar. Furthermore, because CCG is lexicalized, we show it is possible to factor the parsing model over words and introduce a new A∗ parsing algorithm— which we demonstrate is faster and more accurate than adaptive supertagging. Our joint model is the first to substantially im- prove both syntactic and semantic accu- racy over a comparable pipeline, and also achieves state-of-the-art results for a non- ensemble semantic role labelling model. 1 Introduction Joint models of syntactic and semantic parsing are attractive; they can potentially avoid the error propagation that is inherent in pipelines by using semantic models to inform syntactic attachments. However, in practice, the performance of joint sys- tems for semantic role labelling (SRL) has been substantially beneath that of pipelines (Sutton and McCallum, 2005; Lluís et al., 2009; Johansson, 2009; Titov et al., 2009; Naradowsky et al., 2012; Lluís et al., 2013; Henderson et al., 2013). In this paper, we present the first approach to break this trend, by building on the close relationship of syntax and semantics in CCG grammars to enable both (1) a simple but highly effective joint model and (2) an efficient A∗ parsing algorithm. Semantic dependencies can span an unbounded number of syntactic dependencies, causing signif- icant inference and sparsity challenges for joint He refused to confirm or deny the reports nsubj xcomp mark cc conj det dobj ARG0 ARG1 ARG0 ARG1 ARG0 ARG1 Figure 1: Mismatch between syntactic and se- mantic dependencies. He refused to confirm or deny reports NP (S\NP)/(S\NP) S/S (S\NP)/NP conj (S\NP)/NP NP ARG0 ARG1 ∅ ARG0 ARG1 ARG0 ARG1 Figure 2: Dependencies produced by a CCG parse. SRL dependencies can be recovered by la- belling the edges with a semantic role or ∅. Figure 3 shows a CCG derivation for these dependencies. models. For example, in the Figure 1, the se- mantic dependency between He and deny spans three syntactic edges. This fact makes it difficult to jointly parse syntactic and semantic dependen- cies with dynamic programs, and means that de- pendency path features can be sparse. Syntactic dependencies also often have ambiguous seman- tic interpretations—for example in He opened the door and The door opened, the syntactic subject corresponds to different semantic roles. We address these challenges with a new joint model of CCG syntactic parsing and semantic role labelling. The CCG formalism is particu- larly well suited; it …

Content truncated to 500 words