Models and Inference for Prefix-Constrained Machine Translation

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 66–75, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Models and Inference for Prefix-Constrained Machine Translation Joern Wuebker, Spence Green, John DeNero, Saša Hasan Lilt, Inc. first_name@lilt.com Minh-Thang Luong Stanford University lmthang@stanford.edu Abstract We apply phrase-based and neural models to a core task in interactive machine trans- lation: suggesting how to complete a par- tial translation. For the phrase-based sys- tem, we demonstrate improvements in sug- gestion quality using novel objective func- tions, learning techniques, and inference algorithms tailored to this task. Our con- tributions include new tunable metrics, an improved beam search strategy, an n-best extraction method that increases sugges- tion diversity, and a tuning procedure for a hierarchical joint model of alignment and translation. The combination of these tech- niques improves next-word suggestion accu- racy dramatically from 28.5% to 41.2% in a large-scale English-German experiment. Our recurrent neural translation system in- creases accuracy yet further to 53.0%, but inference is two orders of magnitude slower. Manual error analysis shows the strengths and weaknesses of both approaches. 1 Introduction A core prediction task in interactive machine trans- lation (MT) is to complete a partial translation (Ortiz-Martínez et al., 2009; Koehn et al., 2014). Sentence completion enables interfaces that are richer than basic post-editing of MT output. For example, the translator can receive updated sugges- tions after each word typed (Langlais et al., 2000). However, we show that completing partial trans- lations by naïve constrained decoding—the stan- dard in prior work—yields poor suggestion quality. We describe new phrase-based objective functions, learning techniques, and inference algorithms for the sentence completion task.1 We then compare this improved phrase-based system to a state-of-the- art recurrent neural translation system in large-scale English-German experiments. A system for completing partial translations takes as input a source sentence and a prefix of the target sentence. It predicts a suffix: a sequence of tokens that extends the prefix to form a full sentence. In an interactive setting, the first words of the suffix are critical; these words are the focus of the user’s atten- tion and can typically be appended to the translation with a single keystroke. We introduce a tuning met- ric that scores correctness of the whole suffix, but is particularly sensitive to these first words. Phrase-based inference for this task involves aligning the prefix to the source, then generat- ing the suffix by translating the unaligned words. We describe a beam search strategy and a hi- erarchical joint model of alignment and transla- tion that together improve suggestions dramatically. For English-German news, next-word accuracy in- creases from 28.5% to 41.2%. An interactive MT system could also display mul- tiple suggestions to the user. We describe an algo- rithm for efficiently finding the n-best next words directly following a prefix and their corresponding best suffixes. Our experiments show that this ap- proach to n-best list extraction, combined with our other improvements, increased next-word sugges- tion accuracy of 10-best lists from 33.4% …

Content truncated to 500 words