Annotating and Predicting Non-Restrictive Noun Phrase Modifications

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1256–1265, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Annotating and Predicting Non-Restrictive Noun Phrase Modifications Gabriel Stanovsky Ido Dagan Computer Science Department, Bar-Ilan University gabriel.satanovsky@gmail.com dagan@cs.biu.ac.il Abstract The distinction between restrictive and non-restrictive modification in noun phrases is a well studied subject in linguistics. Automatically identifying non-restrictive modifiers can provide NLP applications with shorter, more salient arguments, which were found beneficial by several recent works. While previous work showed that restrictiveness can be annotated with high agreement, no large scale corpus was created, hindering the development of suitable classification algorithms. In this work we devise a novel crowdsourcing annotation methodology, and an accompanying large scale corpus. Then, we present a robust automated system which identifies non-restrictive modifiers, notably improving over prior methods. 1 Introduction Linguistic literature provides a large body of re- search distinguishing between two types of mod- ifiers within noun phrases: (1) Restrictive modi- fiers, which constitute an integral part of the entity referenced by the NP, e.g., the underlined modifier in “She wore the necklace that her mother gave her”, versus (2) Non-restrictive modifiers, which provide an additional or parenthetical information on an already definite entity, e.g., “The speaker thanked president Obama who just came into the room” (Huddleston et al., 2002; Fabb, 1990; Um- bach, 2006). The distinction between the two types is seman- tic in nature and relies heavily on the context of the NP. Evidently, many syntactic constructions can appear in both restrictive and non-restrictive uses. While the previous examples were of rela- tive clauses, Figure 1 demonstrates this distinction in various other syntactic constructions. Identifying and removing non-restrictive mod- ifiers yields shorter NP arguments, which proved beneficial in many NLP tasks. In the context of abstractive summarization (Ganesan et al., 2010) or sentence compression (Knight and Marcu, 2002), non-restrictive modifiers can be removed to shorten sentences, while restrictive modification should be preserved. Further, recent work in information extraction showed that shorter arguments can be beneficial for downstream tasks. Angeli et al. (2015) built an Open-IE system which focuses on shorter ar- gument spans, and demonstrated its usefulness in a state-of-the-art Knowledge Base Population sys- tem. Stanovsky et al. (2015) compared the per- formance of several off-the-shelf analyzers in dif- ferent semantic tasks. Most relevant to this work is the comparison between Open-IE and Seman- tic Role Labeling (Carreras and M`arquez, 2005). Specifically, they suggest that SRL’s longer argu- ments introduce noise which hurts performance for downstream tasks. Finally, in question answering, omitting non- restrictive modification can assist in providing more concise answers, or in matching between multiple answer occurrences. Despite these benefits, there is currently no con- sistent large scale annotation of restrictiveness, which hinders the development of automatic tools for its classification. In prior art in this field, Dor- nescu et al. (2014) used trained annotators to mark restrictiveness in a large corpus. Although they reached good agreement levels in restrictiveness annotation, their corpus suffered from …

Content truncated to 500 words