Rule-Based Translation with Statistical Phrase-Based Post-Editing

Proceedings of the Second Workshop on Statistical Machine Translation, pages 203–206, Prague, June 2007. c⃝2007 Association for Computational Linguistics Rule-based Translation With Statistical Phrase-based Post-editing Michel Simard, Nicola Ueffing, Pierre Isabelle and Roland Kuhn Interactive Language Technologies Group National Research Council of Canada Gatineau, Canada, K1A 0R6 firstname.lastname@nrc-cnrc.gc.ca Abstract This article describes a machine translation system based on an automatic post-editing strategy: initially translate the input text into the target-language using a rule-based MT system, then automatically post-edit the out- put using a statistical phrase-based system. An implementation of this approach based on the SYSTRAN and PORTAGE MT sys- tems was used in the shared task of the Sec- ond Workshop on Statistical Machine Trans- lation. Experimental results on the test data of the previous campaign are presented. 1 Introduction Simard et al. (2007) have recently shown how a sta- tistical phrase-based machine translation system can be used as an automatic post-editing (APE) layer, on top of a rule-based machine translation system. The motivation for their work is the repetitive nature of the errors typically made by rule-based systems. Given appropriate training material, a statistical MT system can be trained to correct these systematic er- rors, therefore reducing the post-editing effort. The statistical system views the output of the rule-based system as the source language, and reference hu- man translations as the target language. Because the training material for the APE layer will typically be domain-specific, this process can be viewed as a way of automatically adapting a rule-based system to a specific application domain. This approach has been shown experimentally to produce large improvements in performance not only over the baseline rule-based system that it cor- rects, but also over a similar statistical phrase-based MT system used in standalone mode, i.e. translating the “real” source text directly: Simard et al. report a reduction in post-editing effort of up to a third when compared to the input rule-based translation, and as much as 5 BLEU points improvement over the direct SMT approach. These impressive results, however, were obtained in a very specific and somewhat unusual context: the training and test corpora were extracted from a collection of manually post-edited machine trans- lations. The two corpora (one English-to-French, one French-to-English) each contained three paral- lel “views” of the same data: 1) the source language text, 2) a machine translation of that text into the target language, as produced by a commercial rule- based MT system, and 3) the final target-language version of the text, produced by manually post- editing the machine translation. Furthermore, the corpus was very small, at least by SMT standards: 500K words of source-language data in the French- to-English direction, 350K words in the English-to- French. Because of this, the authors were left with two important questions: 1) how would the results scale up to much larger quantities of training data? and 2) are the results related to the dependent nature of the translations, i.e. is the automatic post-editing approach still effective when the machine and hu- man …

Content truncated to 500 words