Using Dependency Order Templates to Improve Generality in Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 1–8, Prague, June 2007. c⃝2007 Association for Computational Linguistics Using Dependency Order Templates to Improve Generality in Translation Arul Menezes and Chris Quirk Microsoft Research One Microsoft Way, Redmond, WA 98052, USA {arulm, chrisq}@microsoft.com Abstract Today's statistical machine translation systems generalize poorly to new domains. Even small shifts can cause precipitous drops in translation quality. Phrasal systems rely heavily, for both reordering and contextual translation, on long phrases that simply fail to match out- of-domain text. Hierarchical systems attempt to generalize these phrases but their learned rules are subject to severe constraints. Syntactic systems can learn lexicalized and unlexicalized rules, but the joint modeling of lexical choice and reordering can narrow the applicability of learned rules. The treelet approach models reordering separately from lexical choice, using a discriminatively trained order model, which allows treelets to apply broadly, and has shown better generalization to new domains, but suffers a factorially large search space. We introduce a new reordering model based on dependency order templates, and show that it outperforms both phrasal and treelet systems on in-domain and out-of-domain text, while limiting the search space. 1 Introduction Modern phrasal SMT systems such as (Koehn et al., 2003) derive much of their power from being able to memorize and use long phrases. Phrases allow for non-compositional translation, local reordering and contextual lexical choice. However the phrases are fully lexicalized, which means they generalize poorly to even slightly out- of-domain text. In an open competition (Koehn & Monz, 2006) systems trained on parliamentary proceedings were tested on text from 'news commentary' web sites, a very slightly different domain. The 9 phrasal systems in the English to Spanish track suffered an absolute drop in BLEU score of between 4.4% and 6.34% (14% to 27% relative). The treelet system of Menezes et al. (2006) fared somewhat better but still suffered an absolute drop of 3.61%. Clearly there is a need for approaches with greater powers of generalization. There are multiple facets to this issue, including handling of unknown words, new senses of known words etc. In this work, we will focus on the issue of reordering, i.e. can we learn how to transform the sentence structure of one language into the sentence structure of another, in a way that is not tied to a specific domain or sub-domains, or indeed, sequences of individual words. An early attempt at greater generality in a purely phrasal setting was the alignment template approach (Och & Ney 2004); newer approaches include formally syntactic (Chiang 2005), and linguistically syntactic approaches (Quirk et al. 2005), (Huang et al. 2006). In the next section, we examine these representative approaches to the reordering problem. 2 Related Work Our discussion of related work will be grounded in the following tiny English to Spanish example, where the training set includes: a very old book un libro más antiguo a book very old1 the old man el hombre viejo the man old it is very important es …

Content truncated to 500 words