Supertagged Phrase-Based Statistical Machine Translation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 288–295, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Supertagged Phrase-Based Statistical Machine Translation Hany Hassan School of Computing, Dublin City University, Dublin 9, Ireland hhasan@computing.dcu.ie Khalil Sima’an Language and Computation, University of Amsterdam, Amsterdam, The Netherlands simaan@science.uva.nl Andy Way School of Computing, Dublin City University, Dublin 9, Ireland away@computing.dcu.ie Abstract Until quite recently, extending Phrase-based Statistical Machine Translation (PBSMT) with syntactic structure caused system per- formance to deteriorate. In this work we show that incorporating lexical syntactic de- scriptions in the form of supertags can yield significantly better PBSMT systems. We de- scribe a novel PBSMT model that integrates supertags into the target language model and the target side of the translation model. Two kinds of supertags are employed: those from Lexicalized Tree-Adjoining Grammar and Combinatory Categorial Grammar. De- spite the differences between these two ap- proaches, the supertaggers give similar im- provements. In addition to supertagging, we also explore the utility of a surface global grammaticality measure based on combina- tory operators. We perform various experi- ments on the Arabic to English NIST 2005 test set addressing issues such as sparseness, scalability and the utility of system subcom- ponents. Our best result (0.4688 BLEU) improves by 6.1% relative to a state-of-the- art PBSMT model, which compares very favourably with the leading systems on the NIST 2005 task. 1 Introduction Within the field of Machine Translation, by far the most dominant paradigm is Phrase-based Statistical Machine Translation (PBSMT) (Koehn et al., 2003; Tillmann & Xia, 2003). However, unlike in rule- and example-based MT, it has proven difficult to date to incorporate linguistic, syntactic knowledge in order to improve translation quality. Only quite recently have (Chiang, 2005) and (Marcu et al., 2006) shown that incorporating some form of syntactic structure could show improvements over a baseline PBSMT system. While (Chiang, 2005) avails of structure which is not linguistically motivated, (Marcu et al., 2006) employ syntactic structure to enrich the en- tries in the phrase table. In this paper we explore a novel approach towards extending a standard PBSMT system with syntactic descriptions: we inject lexical descriptions into both the target side of the phrase translation table and the target language model. Crucially, the kind of lexical descriptions that we employ are those that are com- monly devised within lexicon-driven approaches to linguistic syntax, e.g. Lexicalized Tree-Adjoining Grammar (Joshi & Schabes, 1992; Bangalore & Joshi, 1999) and Combinary Categorial Grammar (Steedman, 2000). In these linguistic approaches, it is assumed that the grammar consists of a very rich lexicon and a tiny, impoverished1 set of combina- tory operators that assemble lexical entries together into parse-trees. The lexical entries consist of syn- tactic constructs (‘supertags’) that describe informa- tion such as the POS tag of the word, its subcatego- rization information and the hierarchy of phrase cat- egories that the word projects upwards. In this work we employ the lexical entries but exchange the al- gebraic combinatory operators …

Content truncated to 500 words