Advancements in Arabic-to-English Hierarchical Machine Translation

Advancements in Arabic-to-English Hierarchical Machine Translation Matthias Huck1 and David Vilar1,2 and Daniel Stein1 and Hermann Ney1 1 Human Language Technology and Pattern 2 DFKI GmbH Recognition Group, RWTH Aachen University Berlin, Germany <surname>@cs.rwth-aachen.de david.vilar@dfki.de Abstract In this paper we study several advanced techniques and models for Arabic-to- English statistical machine translation. We examine how the challenges imposed by this particular language pair and transla- tion direction can be successfully tack- led within the framework of hierarchical phrase-based translation. We extend the state-of-the-art with a novel cross-system and cross-paradigm lightly- supervised training approach. In addition, for following recently developed tech- niques we provide a concise review, an em- pirical evaluation, and an in-depth analy- sis: soft syntactic labels, a discriminative word lexicon model, additional reorder- ings, and shallow rules. We thus bring to- gether complementary methods that previ- ously have only been investigated in iso- lation and mostly on different language pairs. Combinations of the methods yield signifi- cant improvements over a baseline using a usual set of models. The resulting hierar- chical systems perform competitive on the large-scale NIST Arabic-to-English trans- lation task. 1 Introduction Since its introduction in (Chiang, 2005), hierar- chical phrase-based translation has become a stan- dard approach in statistical machine translation. Many additional features and enhancements to the hierarchical paradigm have been proposed or c⃝ 2011 European Association for Machine Translation. adopted from the conventional phrase-based ap- proach, but the effect of the various methods is typically merely evaluated separately. Neither are they compared to each other, nor is it clear whether combining the methods would be beneficial. The aim of the work presented in this pa- per is to explore the effectiveness of a state-of- the-art hierarchical phrase-based system for large- scale Arabic-to-English statistical machine trans- lation (SMT). Within this framework, we inves- tigate the impact of several recently developed methods on the translation performance. Not only do we analyze them separately, but also exam- ine whether their combination further increases the output quality. More specifically, we focus on three models: First, we integrate syntactic information in order to improve the linguistic structure of the transla- tion. Second, we utilize a discriminatively trained extended word lexicon to obtain a better lexical selection based on global source sentence con- text. Third, we introduce a jump model which is based on reordering enhancements to the hierar- chical grammar to allow for more flexibility during the search process. The Arabic-English language pair is known to behave more monotone than other language pairs, e.g. Urdu-English or Chinese-English. In a con- trastive experiment done by Birch et al. (2009), a hierarchical system does not outperform a con- ventional phrase-based system for Arabic-English. On the other hand, a lattice-based hierarchical system (de Gispert et al., 2010) has been the best-performing system at the 2009 NIST Arabic- English evaluation campaign.1 Noticing these 1http://www.itl.nist.gov/iad/mig/tests/ mt/2009/ResultsRelease/currentArabic. html Mikel L. Forcada, Heidi Depraetere, Vincent Vandeghinste (eds.) Proceedings of the 15th Conference of the European Association for Machine Translation, p. 273280 Leuven, Belgium, May 2011 …

Content truncated to 500 words