Adjunct Alignment in Translation Data with an Application to Phrase Based Statistical Machine Translation

Adjunct Alignment in Translation Data with an Application to Phrase-Based Statistical Machine Translation Sophie Arnoult Institute of Logic, Language and Computation (ILLC) University of Amsterdam Amsterdam, the Netherlands s.arnoult@gmail.com Khalil Sima’an Institute of Logic, Language and Computation (ILLC) University of Amsterdam Amsterdam, the Netherlands k.simaan@uva.nl Abstract Enriching statistical models with linguis- tic knowledge has been a major concern in Machine Translation (MT). In mono- lingual data, adjuncts are optional con- stituents contributing secondarily to the meaning of a sentence. One can there- fore hypothesize that this secondary status is preserved in translation, and thus that adjuncts may align consistently with their adjunct translations, suggesting they form optional phrase pairs in parallel corpora. In this paper we verify this hypothesis on French-English translation data, and ex- plore the utility of compiling adjunct-poor data for augmenting the training data of a phrase-based machine translation model. 1 Introduction Phrase-Based Statistical Machine Translation (PB- SMT) (Koehn et al., 2003) exploits symmetrized word alignments (Brown et al., 1993) to form phrase pairs that capture the translation probabil- ities of idiomatic expressions. However, data spar- sity is a major issue for phrase-based systems. It affects longer phrase pairs in particular, which are overestimated by the unsmoothed heuristic counts. Smoothing has been proposed to improve proba- bility estimations in the phrase table (Kuhn et al., 2006; Foster et al., 2006), and minimal phrase pairs to alleviate data sparsity: see the tuples of Schwenk (2007) and the minimal translation units of Quirk and Menezes (2006). In both cases, these new units of translation are utilized in an n-gram translation model that allows to capture contextual dependencies, and their estimates are smoothed. c⃝ 2012 European Association for Machine Translation. Morphology has also been proposed to reduce data sparsity, either by integrating morphologi- cal information into the translation model or as a preprocessing step. For instance, Nießen and Ney (2004) propose hierarchical lexicon models in a German-English system, with feature func- tions to integrate different levels of morphosyntac- tic abstraction, from full word forms to lemmas, whereas El Kholy and Habash (2010) present dif- ferent morphological tokenization models for Ara- bic. In this work we propose to use adjuncts to aug- ment and smooth training data for machine trans- lation. The term ‘adjunct’ is used here to refer to both clausal adjuncts and phrase modifiers, regard- less of the nature of the modified category, e.g., verbal or nominal. As optional constituents that further qualify a complete clause or phrase, ad- juncts can be removed from or added to a mono- lingual sentence without affecting its grammatical- ity. This idea is embodied most visibly in Tree- Adjoining Grammar (Joshi, 1983) where recursion in syntax is factored out into auxiliary lexical ele- ments that modify initial sentences by adjunction operation. Here, we hypothesize that adjuncts, by their secondary semantic status, are likely to be preserved in translation. To our knowledge, this constitutes the first study of adjunct alignment in parallel data, though this idea is related to the Di- rect Correspondence Assumption …

Content truncated to 500 words