Agreement Constraints for Statistical Machine Translation into German

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 217–226, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Agreement Constraints for Statistical Machine Translation into German Philip Williams and Philipp Koehn School of Informatics University of Edinburgh 10 Crichton Street EH8 9AB, UK p.j.williams-2@sms.ed.ac.uk pkoehn@inf.ed.ac.uk Abstract Languages with rich inflectional morphology pose a difficult challenge for statistical ma- chine translation. To address the problem of morphologically inconsistent output, we add unification-based constraints to the target-side of a string-to-tree model. By integrating con- straint evaluation into the decoding process, implausible hypotheses can be penalised or filtered out during search. We use a sim- ple heuristic process to extract agreement con- straints for German and test our approach on an English-German system trained on WMT data, achieving a small improvement in trans- lation accuracy as measured by BLEU. 1 Introduction Historically, most work in statistical machine trans- lation (SMT) has focused on translation into En- glish. Languages with richer inflectional mor- phologies pose additional challenges for translation and conventional SMT approaches tend to perform poorly when either source or target language has rich morphology (Koehn, 2005). For complex source inflection, a successful ap- proach has been to cluster inflectional variants into equivalence classes. This removes information that is redundant for translation and can be performed as a preprocessing step for input to a conventional sur- face form based translation model (Nießen and Ney, 2001; Goldwater and McClosky, 2005; Talbot and Osborne, 2006). For complex target inflection, Minkov et al. (2007) investigate how post- processing can be used to generate inflection for a system that produces uninflected output. Their ap- proach is successfully applied to English-Arabic and English-Russian systems by Toutanova et al. (2008). Another promising line of research involves the direct integration of linguistic information into SMT models. Koehn and Hoang (2007) generalise the phrase-based model’s representation of the word from a string to a vector, allowing additional features such as part-of-speech and morphology to be asso- ciated with, or even to replace, surface forms dur- ing search. Luong et al. (2010) decompose words into morphemes and use this extended represen- tation throughout the training, tuning, and testing pipeline. Departing further from traditional SMT mod- els, the transfer-based systems of Riezler and Maxwell (2006), Bojar and Hajiˇc (2008), and Gra- ham et al. (2009) employ rich feature structure representations for linguistic attributes, but have so far been limited by their dependence on non- stochastic parsers with limited coverage. The Stat- XFER transfer-based framework (Lavie, 2008) is neutral with regard to the rule acquisition method and the author describes a manually developed Hebrew-English transfer grammar, which includes a small number of constraints between agreement fea- tures. In Hanneman et al. (2009) the framework is used with a large automatically-extracted grammar, though this does not use feature constraints. In this paper we propose a model that retains the use of surface forms during decoding whilst also checking linguistic constraints defined over asso- ciated feature structures. Specifically, we extend a string-to-tree …

Content truncated to 500 words