Universal Decompositional Semantics on Universal Dependencies

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1713–1723, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Universal Decompositional Semantics on Universal Dependencies Aaron Steven White Drew Reisinger Keisuke Sakaguchi Tim Vieira Sheng Zhang Rachel Rudinger Kyle Rawlins Benjamin Van Durme Johns Hopkins University Abstract We present a framework for augmenting data sets from the Universal Dependencies project with Universal Decompositional Semantics. Where the Universal Dependencies project aims to provide a syntactic annotation stan- dard that can be used consistently across many languages as well as a collection of corpora that use that standard, our extension has simi- lar aims for semantic annotation. We describe results from annotating the English Univer- sal Dependencies treebank, dealing with word senses, semantic roles, and event properties. 1 Introduction This paper describes the Universal Decompositional Semantics (Decomp) project, which aims to aug- ments Universal Dependencies (UD) data sets with robust, scalable semantic annotations based in lin- guistic theory. The UD project1 aims to provide (i) a syntactic dependency annotation standard that can be used consistently across many languages and (ii) a collection of corpora that use that stan- dard (De Marneffe et al., 2014; Nivre et al., 2015). Decomp provides complementary semantic annota- tions that scale across different types of semantic in- formation and different languages and can integrate seamlessly with any UD-annotated corpus. Decomp has two mutually supportive tenets— semantic decomposition and simplicity. As we dis- cuss further in the next section, these tenets allow us to collect annotations from everyday speakers of a language that are rooted in basic, commonsensical 1http://universaldependencies.org Your buddy Beau invited me amod nmod:poss nsubj dobj AWARENESS+ VOLITION + ... ... ARG ARG ...REALIS + ... ...ANIMATE + ... Figure 1: Decompositional semantics atop syntax. aspects of meaning and that can be straightforwardly explained and generally agreed upon in context. In this paper, we describe Decomp protocols for three domains—semantic role decomposition, event decomposition, and word sense decomposition— and we present annotation results on top of the En- glish UD v1.2 (EUD1.2) treebank.2 We begin in §2 by connecting Decomp with previous work on de- composition in linguistic theory. In §3, we present PredPatt, which is a software package for prepro- cessing UD annotated corpora for input into De- comp protocols. In §4, we present a major revi- sion to Reisinger et al.’s (2015) semantic role de- composition protocol (SPR1). Our revision, SPR2, brings SPR1 into full alignment with Decomp while adding various new features. In §5, we present Decomp-aligned annotation of event properties, fo- cusing specifically on event realis. Finally, in §6, we describe a Decomp-aligned word sense decomposi- tion protocol and associated set of annotations. 2 Universal Decompositional Semantics A range of perspectives suggest that the proper rep- resentation for word meanings is decompositional. 2All datasets are available at http://decomp.net. 1713 For example, Dowty (1979) followed by a sub- stantial amount of research (e.g. Jackendoff 1990; Rappaport-Hovav and Levin 1998; Levin and Rap- paport Hovav 2005) suggests that word …

Content truncated to 500 words