Model With Minimal Translation Units, But Decode With Phrases

Proceedings of NAACL-HLT 2013, pages 1–11, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Model With Minimal Translation Units, But Decode With Phrases Nadir Durrani∗ University of Edinburgh dnadir@inf.ed.ac.uk Alexander Fraser Helmut Schmid University of Stuttgart fraser,schmid@ims.uni-stuttgart.de Abstract N-gram-based models co-exist with their phrase-based counterparts as an alternative SMT framework. Both techniques have pros and cons. While the N-gram-based frame- work provides a better model that captures both source and target contexts and avoids spurious phrasal segmentation, the ability to memorize and produce larger translation units gives an edge to the phrase-based systems dur- ing decoding, in terms of better search per- formance and superior selection of transla- tion units. In this paper we combine N-gram- based modeling with phrase-based decoding, and obtain the benefits of both approaches. Our experiments show that using this combi- nation not only improves the search accuracy of the N-gram model but that it also improves the BLEU scores. Our system outperforms state-of-the-art phrase-based systems (Moses and Phrasal) and N-gram-based systems by a significant margin on German, French and Spanish to English translation tasks. 1 Introduction Statistical Machine Translation advanced from word-based models (Brown et al., 1993) towards more sophisticated models that take contextual in- formation into account. Phrase-based (Och and Ney, 2004; Koehn et al., 2003) and N-gram-based (Mari˜no et al., 2006) models are two instances of such frameworks. While the two models have some common properties, they are substantially different. ∗Much of the work presented here was carried out while the first author was at the University of Stuttgart. Phrase-based systems employ a simple and effec- tive machinery by learning larger chunks of trans- lation called phrases1. Memorizing larger units en- ables the phrase-based model to learn local depen- dencies such as short reorderings, idioms, insertions and deletions, etc. The model however, has the fol- lowing drawbacks: i) it makes independence as- sumptions over phrases ignoring the contextual in- formation outside of phrases ii) it has issues han- dling long-distance reordering iii) it has the spurious phrasal segmentation problem which allows multi- ple derivations of a bilingual sentence pair having different model scores for each segmentation. Modeling with minimal translation units helps ad- dress some of these issues. The N-gram-based SMT framework is based on tuples. Tuples are mini- mal translation units composed of source and target cepts2. N-gram-based models are Markov models over sequences of tuples (Mari˜no et al., 2006; Crego and Mari˜no, 2006) or operations encapsulating tu- ples (Durrani et al., 2011). This mechanism has sev- eral useful properties. Firstly, no phrasal indepen- dence assumption is made. The model has access to both source and target context outside of phrases. Secondly the model learns a unique derivation of a bilingual sentence given its alignment, thus avoiding the spurious segmentation problem. Using minimal translation units, however, results in a higher number of search errors because of i) 1A phrase-pair in PBSMT is a sequence of source and target words that is translation of each other, and is not necessarily a …

Content truncated to 500 words