Chunk-Based Verb Reordering in VSO Sentences for Arabic-English Statistical Machine Translation

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 235–243, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Chunk-based Verb Reordering in VSO Sentences for Arabic-English Statistical Machine Translation Arianna Bisazza and Marcello Federico Fondazione Bruno Kessler Human Language Technologies Trento, Italy {bisazza,federico}@fbk.eu Abstract In Arabic-to-English phrase-based statis- tical machine translation, a large number of syntactic disfluencies are due to wrong long-range reordering of the verb in VSO sentences, where the verb is anticipated with respect to the English word order. In this paper, we propose a chunk-based reordering technique to automatically de- tect and displace clause-initial verbs in the Arabic side of a word-aligned parallel cor- pus. This method is applied to preprocess the training data, and to collect statistics about verb movements. From this anal- ysis, specific verb reordering lattices are then built on the test sentences before de- coding them. The application of our re- ordering methods on the training and test sets results in consistent BLEU score im- provements on the NIST-MT 2009 Arabic- English benchmark. 1 Introduction Shortcomings of phrase-based statistical machine translation (PSMT) with respect to word reorder- ing have been recently shown on the Arabic- English pair by Birch et al. (2009). An empiri- cal investigation of the output of a strong baseline we developed with the Moses toolkit (Koehn et al., 2007) for the NIST 2009 evaluation, revealed that an evident cause of syntactic disfluency is the anticipation of the verb in Arabic Verb-Subject- Object (VSO) sentences – a class that is highly represented in the news genre1. Fig. 1 shows two examples where the Arabic main verb phrase comes before the subject. In such sentences, the subject can be followed by adjectives, adverbs, coordinations, or appositions that further increase the distance between the verb 1In fact, Arabic syntax admits both SVO and VSO orders. and its object. When translating into English – a primarily SVO language – the resulting long verb reorderings are often missed by the PSMT decoder either because of pure modeling errors or because of search errors (Germann et al., 2001): i.e. their span is longer than the maximum allowed distor- tion distance, or the correct reordering hypothesis does not emerge from the explored search space because of a low score. In the two examples, the missed verb reorderings result in different transla- tion errors by the decoder, respectively, the intro- duction of a subject pronoun before the verb and, even worse, a verbless sentence. In Arabic-English machine translation, other kinds of reordering are of course very frequent: for instance, adjectival modifiers following their noun and head-initial genitive constructions (Idafa). These, however, appear to be mostly local, there- fore more likely to be modeled through phrase in- ternal alignments, or to be captured by the reorder- ing capabilities of the decoder. In general there is a quite uneven distribution of word-reordering phe- nomena in Arabic-English, and long-range move- ments concentrate on few patterns. Reordering in PSMT is typically performed by (i) constraining …

Content truncated to 500 words