Content Differences in Syntactic and Semantic Representation

Proceedings of NAACL-HLT 2019, pages 478–488 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 478 Content Differences in Syntactic and Semantic Representations Daniel Hershcovich1,2 Omri Abend2 1The Edmond and Lily Safra Center for Brain Sciences 2School of Computer Science and Engineering Hebrew University of Jerusalem {danielh,oabend,arir}@cs.huji.ac.il Ari Rappoport2 Abstract Syntactic analysis plays an important role in semantic parsing, but the nature of this role re- mains a topic of ongoing debate. The debate has been constrained by the scarcity of empiri- cal comparative studies between syntactic and semantic schemes, which hinders the develop- ment of parsing methods informed by the de- tails of target schemes and constructions. We target this gap, and take Universal Dependen- cies (UD) and UCCA as a test case. After abstracting away from differences of conven- tion or formalism, we find that most content divergences can be ascribed to: (1) UCCA’s distinction between a Scene and a non-Scene; (2) UCCA’s distinction between primary re- lations, secondary ones and participants; (3) different treatment of multi-word expressions, and (4) different treatment of inter-clause link- age. We further discuss the long tail of cases where the two schemes take markedly differ- ent approaches. Finally, we show that the pro- posed comparison methodology can be used for fine-grained evaluation of UCCA pars- ing, highlighting both challenges and poten- tial sources for improvement. The substantial differences between the schemes suggest that semantic parsers are likely to benefit down- stream text understanding applications beyond their syntactic counterparts. 1 Introduction Semantic representations hold promise due to their ability to transparently reflect distinctions relevant for text understanding applications. For example, syntactic representations are usually sensitive to distinctions based on POS (part of speech), such as between compounds and possessives. Seman- tic schemes are less likely to make this distinc- tion since a possessive can often be paraphrased as a compound and vice versa (e.g., “US presi- dent”/“president of the US”), but may distinguish different senses of possessives (e.g., “some of the presidents” and “inauguration of the presidents”). Nevertheless, little empirical study has been done on what distinguishes semantic schemes from syntactic ones, which are still in many cases the backbone of text understanding sys- tems. Such studies are essential for (1) determin- ing whether and to what extent semantic methods should be adopted for text understanding applica- tions; (2) defining better inductive biases for se- mantic parsers, and allowing better use of infor- mation encoded in syntax; (3) pointing at semantic distinctions unlikely to be resolved by syntax. The importance of such an empirical study is emphasized by the ongoing discussion as to what role syntax should play in semantic parsing, if any (Swayamdipta et al., 2018; Strubell et al., 2018; He et al., 2018; Cai et al., 2018). See §8. This paper aims to address this gap, focusing on content differences. As a test case, we compare relatively similar schemes (§2): the syntactic Uni- versal Dependencies (UD; Nivre et al., 2016), and the semantic Universal Conceptual Cognitive An- …

Content truncated to 500 words