Multiword Expressions in the Context of Statistical Machine Translation

International Joint Conference on Natural Language Processing, pages 1181–1187, Nagoya, Japan, 14-18 October 2013. Multiword Expressions in the Context of Statistical Machine Translation Mahmoud Ghoneim Mona Diab Center for Computational Learning Systems Department of Computer Science Columbia University George Washington University 475 Riverside Drive MC 7717 801 22nd Street NW New York, NY 10115 Washington DC 20052 mah.ghoneim@gmail.com mtdiab@email.gwu.edu Abstract Incorporating semantic information in the Statistical Machine Translation (SMT) framework is starting to gain some populari- ty in both the semantics and translation communities. In this paper, we present en- couraging results obtained from experi- ments conducted on English to Arabic SMT system using static, dynamic, and hybrid in- tegration of fine-grained Multiword Expres- sion (MWE). We achieve an improvement up to 0.82 absolute BLEU score by integrat- ing MWEs over a vanilla SMT system. We empirically show that different MWE types require different integration methods in the SMT framework. 1 Introduction Multiword expressions (MWEs) are roughly de- fined by (Sag et al., 2002) as “idiosyncratic con- cepts that cross word boundaries (spaces).” MWEs are widely used, 41% of the entries in WordNet 1.7 (Fellbaum, 1998) are MWEs, but unfortunately they have proved to be hard to model in natural language processing applica- tions. Typical statistical machine translation (SMT) systems, in particular, do not explicitly model MWEs. This might indicate that state of the art SMT systems are doing well without hav- ing any knowledge of whether a given phrase is a multiword expression or not. However, recent research (Carpuat and Diab 2010, Bouamor et al., 2012) show that explicitly modeling MWEs in the SMT framework yields non-negligible gains depending on the integration method. In this paper we study explicit modeling of the diverse kinds of MWEs in a phrase-based SMT framework for the English-Arabic language pair. This paper is organized as follows: section 2 overviews the different types of MWEs, section 3 reviews the previous work related to MWEs and SMT. Section 4 details our approach fol- lowed by the results in section 5. Our discussion of the results is presented in section 6 and finally the conclusions are in section 7. 2 Multiword Expressions Classification According to (Sag et al., 2002), MWEs are broadly classified into institutionalized phrases and lexicalized phrases based on the varying de- gree of lexical rigidity and semantic composi- tionality. Institutionalized phrases are conventional- ized phrases that are syntactically and semanti- cally compositional, but statistically idiosyncratic (e.g. “traffic light”, “to kindle ex- citement”.) Lexicalized phrases have at least in part idio- syncratic syntax or semantics. They can be fur- ther broken down into: (a) Fixed expressions which undergo neither morphosyntactic variation, nor internal modifica- tion (e.g. “by and large”, “every which way”) [AV, AJ], (b) Semi-fixed expressions such as (1) non- decomposable idioms (e.g. “kick the bucket”) [VNC], (2) compound nominal (e.g. “car park”, “part of speech”) [NNC], and (3) proper names and named entities (e.g. “New York”) [NE]. (c) Syntactically-flexible expressions such as (1) verb particle construction (e.g. “write up”, “look up”) [VPC], …

Content truncated to 500 words