CCG Syntactic Reordering Models for Phrase-based Machine Translation

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 210–221, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics CCG Syntactic Reordering Models for Phrase-based Machine Translation Dennis N. Mehay The Ohio State University Columbus, OH, USA mehay@ling.ohio-state.edu Chris Brew Educational Testing Service Princeton, NJ, USA cbrew@ets.org Abstract Statistical phrase-based machine translation requires no linguistic information beyond word-aligned parallel corpora (Zens et al., 2002; Koehn et al., 2003). Unfortunately, this linguistic agnosticism often produces un- grammatical translations. Syntax, or sentence structure, could provide guidance to phrase- based systems, but the “non-constituent” word strings that phrase-based decoders manipu- late complicate the use of most recursive syn- tactic tools. We address these issues by using Combinatory Categorial Grammar, or CCG, (Steedman, 2000), which has a much more flexible notion of constituency, thereby providing more labels for putative non- constituent multiword translation phrases. Us- ing CCG parse charts, we train a syntactic analogue of a lexicalized reordering model by labelling phrase table entries with multiword labels and demonstrate significant improve- ments in translating between Urdu and En- glish, two language pairs with divergent sen- tence structure. 1 Introduction Statistical phrase-based machine translation (PMT) is attractive, as it requires no linguistic informa- tion beyond word-aligned parallel corpora (Zens et al., 2002; Koehn et al., 2003). Unfortunately, this linguistic agnosticism leaves phrase-based systems with no precise characterization of the word order relationships between languages, often leading to ungrammatical translations. Syntax could provide guidance to phrase-based systems, by steering them towards reorderings that reflect the structural rela- tionships between languages, but using syntax to guide a phrase-based system is problematic. Phrase- based systems build the result incrementally from the beginning of the target string to the end, and the intermediate strings need not constitute complete traditional syntactic constituents. It is difficult to reconcile traditional recursive syntactic processing with this regime, because not all intermediate strings considered by the decoder would even have a syntac- tic category to assess. As a result, most phrase-based decoders control reordering using simple distance- based distortion models, which penalize all reorder- ing equally, and lexicalized reordering models (Till- mann, 2004; Axelrod et al., 2005), which probabilis- tically score various reordering configurations con- ditioned on specific lexical translations. While un- doubtedly better than nothing, these models perform poorly when languages diverge considerably in sen- tence structure. Distance-based distortion models are too coarse-grained to distinguish correct from incorrect reordering, while lexical reordering mod- els suffer from data sparsity and fail to capture more general patterns. We argue that finding a way to label translation phrases with syntactic labels will abstract over the observed reordering configurations thereby address both all three deficiencies of granu- larity, data sparsity and lack of generality. The present work presents a novel syntactic ana- logue of the lexicalized reordering model that uses multiword syntactic labels to capture the general re- ordering patterns between two languages with very different word order. We accomplish this by using Combinatory Categorial Grammar, or CCG (Steed- 210 man, 2000), a word-centered syntax …

Content truncated to 500 words