Generating English Determiners in Phrase-Based Translation with Synthetic Translation Options

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 271–280, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Generating English Determiners in Phrase-Based Translation with Synthetic Translation Options Yulia Tsvetkov Chris Dyer Lori Levin Language Technologies InstituteArchna Bhatia Carnegie Mellon University Pittspurgh, PA, 15213, USA {ytsvetko, cdyer, lsl, archna}@cs.cmu.edu Abstract We propose a technique for improving the quality of phrase-based translation systems by creating synthetic translation options—phrasal translations that are gen- erated by auxiliary translation and post- editing processes—to augment the de- fault phrase inventory learned from par- allel data. We apply our technique to the problem of producing English deter- miners when translating from Russian and Czech, languages that lack definiteness morphemes. Our approach augments the English side of the phrase table using a classifier to predict where English arti- cles might plausibly be added or removed, and then we decode as usual. Doing so, we obtain significant improvements in quality relative to a standard phrase-based baseline and to a to post-editing complete translations with the classifier. 1 Introduction Phrase-based translation works as follows. A set of candidate translations for an input sentence is created by matching contiguous spans of the in- put against an inventory of phrasal translations, reordering them into a target-language appropri- ate order, and choosing the best one according to a discriminative model that combines features of the phrases used, reordering patterns, and target lan- guage model (Koehn et al., 2003). This relatively simple approach to translation can be remarkably effective, and, since its introduction, it has been the basis for further innovations, including devel- oping better models for distinguishing the good translations from bad ones (Chiang, 2012; Gim- pel and Smith, 2012; Cherry and Foster, 2012; Eidelman et al., 2013), improving the identifica- tion of phrase pairs in parallel data (DeNero et al., 2008; DeNero and Klein, 2010), and formal gen- eralizations to gapped rules and rich nonterminal types (Chiang, 2007; Galley et al., 2006). This paper proposes a different mechanism for improv- ing phrase-based translation: the use of synthetic translation options to supplement the standard phrasal inventory used in phrase-based translation systems. In the following, we argue that phrase tables ac- quired in usual way will be expected to have gaps in their coverage in certain language pairs and that supplementing these with synthetic translation options is a priori preferable to alternative tech- niques, such as post processing, for generalizing beyond the translation pairs observable in training data (§2). As a case study, we consider the prob- lem of producing English definite/indefinite arti- cles (the, a, and an) when translating from Russian and Czech, two languages that lack overt definite- ness morphemes (§3). We develop a classifier that predicts the presence and absence of English arti- cles (§4). This classifier is used to generate syn- thetic translation options that are used to augment phrase tables used the usual way (§5). We eval- uate their performance relative to post-processing approach and to a baseline phrase-based system, finding that synthetic translation options …

Content truncated to 500 words