Machine Learning for Hybrid Machine Translation

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 312–316, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Machine Learning for Hybrid Machine Translation Sabine Hunsicker DFKI GmbH Language Technology Lab Saarbr¨ucken, Germany sabine.hunsicker@dfki.de Chen Yu DFKI GmbH Language Technology Lab Saarbr¨ucken, Germany yu.chen@dfki.de Christian Federmann DFKI GmbH Language Technology Lab Saarbr¨ucken, Germany cfedermann@dfki.de Abstract We describe a substitution-based system for hybrid machine translation (MT) that has been extended with machine learning components controlling its phrase selection. The approach is based on a rule-based MT (RBMT) system which creates template translations. Based on the rule-based generation parse tree and target-to-target alignments, we identify the set of “interesting” translation candidates from one or more translation engines which could be substituted into our translation templates. The substitution process is either controlled by the output from a binary classifier trained on feature vectors from the different MT engines, or it is depending on weights for the decision factors, which have been tuned using MERT. We are able to observe improvements in terms of BLEU scores over a baseline version of the hybrid system. 1 Introduction In recent years, machine translation (MT) systems have achieved increasingly better translation quality. Still each paradigm has its own challenges: while statistical MT (SMT) systems suffer from a lack of grammatical structure, resulting in ungrammatical sentences, RBMT systems have to deal with a lack of lexical coverage. Hybrid architectures intend to combine the advantages of the individual paradigms to achieve an overall better translation. Federmann et al. (2010) and Federmann and Hun- sicker (2011) have shown that using a substitution- based approach can improve the translation quality of a baseline RBMT system. Our submission to WMT12 is a new, improved version following these approaches. The output of an RBMT engine serves as our translation backbone, and we substitute noun phrases by translations mined from other systems. 2 System Architecture Our hybrid MT system combines translation output from: a) the Lucy RBMT system, described in more detail in (Alonso and Thurmair, 2003); b) the Linguatec RBMT system (Aleksic and Thurmair, 2011); c) Moses (Koehn et al., 2007); d) Joshua (Li et al., 2009). Lucy provides us with the translation skeleton, which is described in more detail in Section 2.2 while systems b)–d) are aligned to this translation template and mined for substitution candidates. We give more detailed information on these systems in Section 2.3. 2.1 Basic Approach We first identify “interesting” phrases inside the rule-based translation and then compute the most probable correspondences in the translation output from the other systems. For the resulting phrases, we apply a factored substitution method that decides whether the original RBMT phrase should be kept or rather be replaced by one of the candidate phrases. A schematic overview of our hybrid system and its main components is given in Figure 1. 312 Figure 1: Schematic overview of the architecture of our substitution-based, hybrid MT system. In previous years, it turned out that the alignment of the candidate translations to the …

Content truncated to 500 words