Applying Pairwise Ranked Optimisation to Improve the Interpolation of Translation Models

Proceedings of NAACL-HLT 2013, pages 342–347, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Applying Pairwise Ranked Optimisation to Improve the Interpolation of Translation Models Barry Haddow University of Edinburgh Scotland bhaddow@inf.ed.ac.uk Abstract In Statistical Machine Translation we often have to combine different sources of parallel training data to build a good system. One way of doing this is to build separate translation models from each data set and linearly inter- polate them, and to date the main method for optimising the interpolation weights is to min- imise the model perplexity on a heldout set. In this work, rather than optimising for this indi- rect measure, we directly optimise for BLEU on the tuning set and show improvements in average performance over two data sets and 8 language pairs. 1 Introduction Statistical Machine Translation (SMT) requires large quantities of parallel training data in order to produce high quality translation systems. This train- ing data, however, is often scarce and must be drawn from whatever sources are available. If these data sources differ systematically from each other, and/or from the test data, then the problem of combining these disparate data sets to create the best possible translation system is known as domain adaptation. One approach to domain adaptation is to build separate models for each training domain, then weight them to create a system tuned to the test do- main. In SMT, a successful approach to building do- main specific language models is to build one from each corpus, then linearly interpolate them, choos- ing weights that minimise the perplexity on a suit- able heldout set of in-domain data. This method has been applied by many authors (e.g. (Koehn and Schroeder, 2007)), and is implemented in popular language modelling tools like IRSTLM (Federico et al., 2008) and SRILM (Stolcke, 2002). Similar interpolation techniques have been devel- oped for translation model interpolation (Foster et al., 2010; Sennrich, 2012) for phrase-based systems but have not been as widely adopted, perhaps be- cause the efficacy of the methods is not as clear- cut. In this previous work, the authors used stan- dard phrase extraction heuristics to extract phrases from a heldout set of parallel sentences, then tuned the translation model (i.e. the phrase table) inter- polation weights to minimise the perplexity of the interpolated model on this set of extracted phrases. In this paper, we try to improve on this perplexity optimisation of phrase table interpolation weights by addressing two of its shortcomings. The first prob- lem is that the perplexity is not well defined because of the differing coverage of the phrase tables, and their partial coverage of the phrases extracted from the heldout set. Secondly, perplexity may not corre- late with the performance of the final SMT system. So, instead of optimising the interpolation weights for the indirect goal of translation model perplexity, we optimise them directly for transla- tion performance. We do this by incorporating these weights into SMT tuning using a modified version of Pairwise Ranked Optimisation (PRO) …

Content truncated to 500 words