Intersecting Multilingual Data for Faster and Better Statistical Translations

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 128–136, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Intersecting multilingual data for faster and better statistical translations Yu Chen1,2, Martin Kay1,3, Andreas Eisele1,2 1: Universit¨at des Saarlandes, Saarbr¨ucken, Germany 2: Deutsches Forschungszentrum f¨ur K¨unstliche Intelligenz GmbH, Saarbr¨ucken, Germany 3: Stanford University, CA, USA {yuchen,kay,eisele}@coli.uni-saarland.de Abstract In current phrase-based SMT systems, more training data is generally better than less. However, a larger data set eventually intro- duces a larger model that enlarges the search space for the translation problem, and con- sequently requires more time and more re- sources to translate. We argue redundant in- formation in a SMT system may not only de- lay the computations but also affect the qual- ity of the outputs. This paper proposes an ap- proach to reduce the model size by filtering out the less probable entries based on com- patible data in an intermediate language, a novel use of triangulation, without sacrificing the translation quality. Comprehensive exper- iments were conducted on standard data sets. We achieved significant quality improvements (up to 2.3 BLEU points) while translating with reduced models. In addition, we demon- strate a straightforward combination method for more progressive filtering. The reduction of the model size can be up to 94% with the translation quality being preserved. 1 Introduction Statistical machine translation (SMT) applies ma- chine learning techniques to a bilingual corpus to produce a translation system entirely automatically. Such a scheme has many potential advantages over earlier systems which relied on carefully crafted rules. The most obvious is that it at dramatically reduces cost in human labor and it is able to reach many critical translation rules that are easily over- looked by human being. SMT systems generally assemble translations by selecting phrases from a large candidate set. Un- supervised learning often introduces a considerable amount of noise into this set as a result of which the selection process becomes more longer and less ef- fective. This paper provides one approach to these problems. Various filtering techniques, such as (Johnson et al., 2007) and (Chen et al., 2008), have been ap- plied to eliminate a large portion of the translation rules that were judged unlikely to be of value for the current translation. However, these approaches were only able to improve the translation quality slightly. In this paper, we describe a triangulation approach (Kay, 1997) that incorporates multilingual data to improve system efficiency and translation quality at the same time. Most of the previous tri- angulation approaches (Kumar et al., 2007; Cohn and Lapata, 2007; Filali and Bilmes, 2005; Simard, 1999; Och and Ney, 2001) add information obtained from a third language. In other words, they work with the union of the data from the different lan- guages. In contrast, we work with the intersection of information acquired through a third language. The hope is that the intersection will be more precise and more compact than the union, so that a better …

Content truncated to 500 words