Multiword Expression Identification with Tree Substitution Grammars: A Parsing tour de force with French

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 725–735, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Multiword Expression Identification with Tree Substitution Grammars: A Parsing tour de force with French Spence Green*, Marie-Catherine de Marneffe†, John Bauer*, and Christopher D. Manning*† *Computer Science Department, Stanford University †Linguistics Department, Stanford University {spenceg,mcdm,horatio,manning}@stanford.edu Abstract Multiword expressions (MWE), a known nui- sance for both linguistics and NLP, blur the lines between syntax and semantics. Previous work on MWE identification has relied primar- ily on surface statistics, which perform poorly for longer MWEs and cannot model discontin- uous expressions. To address these problems, we show that even the simplest parsing mod- els can effectively identify MWEs of arbitrary length, and that Tree Substitution Grammars achieve the best results. Our experiments show a 36.4% F1 absolute improvement for French over an n-gram surface statistics baseline, cur- rently the predominant method for MWE iden- tification. Our models are useful for several NLP tasks in which MWE pre-grouping has improved accuracy. 1 Introduction Multiword expressions (MWE) have long been a challenge for linguistic theory and NLP. There is no universally accepted definition of the term, but MWEs can be characterized as “idiosyncratic inter- pretations that cross word boundaries (or spaces)” (Sag et al., 2002) such as traffic light, or as “fre- quently occurring phrasal units which are subject to a certain level of semantic opaqueness, or non- compositionality” (Rayson et al., 2010). MWEs are often opaque fixed expressions, al- though the degree to which they are fixed can vary. Some MWEs do not allow morphosyntactic varia- tion or internal modification (e.g., in short, but *in shorter or *in very short). Other MWEs are “semi- fixed,” meaning that they can be inflected or undergo internal modification. The type of modification is of- ten limited, but not predictable, so it is not possible to enumerate all variants (Table 1). French English à terme in the near term à court terme in the short term à très court terme in the very short term à moyen terme in the mediumterm à long terme in the long term à très long terme in the very long term Table 1: Semi-fixed MWEs in French and English. The French adverb à terme ‘in the end’ can be modified by a small set of adjectives, and in turn some of these ad- jectives can be modified by an adverb such as très ‘very’. Similar restrictions appear in English. Merging known MWEs into single tokens has been shown to improve accuracy for a variety of NLP tasks: dependency parsing (Nivre and Nilsson, 2004), constituency parsing (Arun and Keller, 2005), sentence generation (Hogan et al., 2007), and ma- chine translation (Carpuat and Diab, 2010). Most ex- periments use gold MWE pre-grouping or language- specific resources like WordNet. For unlabeled text, the best MWE identification methods, which are based on surface statistics (Pecina, 2010), suffer from sparsity induced by longer n-grams (Ramisch et al., 2010). A dilemma thus …

Content truncated to 500 words