Confidence-based Rewriting of Machine Translation Output

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1261–1272, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Confidence-based Rewriting of Machine Translation Output Benjamin Marie LIMSI-CNRS, Orsay, France Lingua et Machina, Le Chesnay, France benjamin.marie@limsi.fr Aur´elien Max LIMSI-CNRS, Orsay, France Univ. Paris Sud, Orsay, France aurelien.max@limsi.fr Abstract Numerous works in Statistical Machine Translation (SMT) have attempted to iden- tify better translation hypotheses obtained by an initial decoding using an improved, but more costly scoring function. In this work, we introduce an approach that takes the hypotheses produced by a state-of- the-art, reranked phrase-based SMT sys- tem, and explores new parts of the search space by applying rewriting rules se- lected on the basis of posterior phrase- level confidence. In the medical do- main, we obtain a 1.9 BLEU improve- ment over a reranked baseline exploiting the same scoring function, corresponding to a 5.4 BLEU improvement over the orig- inal Moses baseline. We show that if an indication of which phrases require rewrit- ing is provided, our automatic rewriting procedure yields an additional improve- ment of 1.5 BLEU. Various analyses, in- cluding a manual error analysis, further il- lustrate the good performance and poten- tial for improvement of our approach in spite of its simplicity. 1 Introduction The standard configuration of modern phrase- based Statistical Machine Translation (SMT) (Koehn et al., 2003) systems can produce very ac- ceptable results on some tasks. However, early integration of better features to guide the search for the best hypothesis can result in significant im- provements, an expression of the complexity of modeling translation quality. For instance, im- provements have been obtained by integrating fea- tures into decoding that better model semantic co- herence at the sentence level (Hasan and Ney, 2009) or syntactic well-formedness (Schwartz et al., 2011). However, early use of such complex features typically comes at a high computational cost. Moreover, some informative features require or are better computed when complete translation hypotheses are available. This is addressed in nu- merous works on reranking of the highest scored sub-space of hypotheses, on so-called n-best lists (Och et al., 2004; Zhang et al., 2006; Carter and Monz, 2011) or output lattices (Schwenk et al., 2006; Blackwood et al., 2010), where many works specifically target the inclusion of better language modelling capabilities, a well-known weakness of current automatic generation approaches (Knight, 2007). Another way to improve translation a posteriori can be done by rewriting initial hypotheses, for in- stance in a greedy fashion by including new mod- els (Langlais et al., 2007; Hardmeier et al., 2012), or by specifically modeling a task of automatic post-editing targeting a specific system (Simard et al., 2007; Dugast et al., 2007). While such auto- matic post-editing may seem to be too limited, no- tably because of the limited initial diversity con- sidered and the fact that it may be in some in- stances agnostic to the internals of the initial sys- tem, it has been shown to potentially improve ac- …

Content truncated to 500 words