Language and Translation Model Adaptation using Comparable Corpora

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 857–866, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Language and Translation Model Adaptation using Comparable Corpora Matthew Snover and Bonnie Dorr Laboratory for Computational Linguistics and Information Processing Institute for Advanced Computer Studies Department of Computer Science University of Maryland College Park, MD 20742 {snover, bonnie}@umiacs.umd.edu Richard Schwartz BBN Technologies 10 Moulton Street Cambridge, MA 02138, USA schwartz@bbn.com Abstract Traditionally, statistical machine translation systems have relied on parallel bi-lingual data to train a translation model. While bi-lingual parallel data are expensive to generate, mono- lingual data are relatively common. Yet mono- lingual data have been under-utilized, having been used primarily for training a language model in the target language. This paper de- scribes a novel method for utilizing monolin- gual target data to improve the performance of a statistical machine translation system on news stories. The method exploits the exis- tence of comparable text—multiple texts in the target language that discuss the same or similar stories as found in the source language document. For every source document that is to be translated, a large monolingual data set in the target language is searched for docu- ments that might be comparable to the source documents. These documents are then used to adapt the MT system to increase the prob- ability of generating texts that resemble the comparable document. Experimental results obtained by adapting both the language and translation models show substantial gains over the baseline system. 1 Introduction While the amount of parallel data available to train a statistical machine translation system is sharply lim- ited, vast amounts of monolingual data are generally available, especially when translating to languages such as English. Yet monolingual data are generally only used to train the language model of the trans- lation system. Previous work (Fung and Yee, 1998; Rapp, 1999) has sought to learn new translations for words by looking at comparable, but not parallel, corpora in multiple languages and analyzing the co- occurrence of words, resulting in the generation of new word-to-word translations. More recently, Resnik and Smith (2003) and Munteanu and Marcu (2005) have exploited monolingual data in both the source and target languages to find document or sentence pairs that appear to be parallel. This newly discovered bilin- gual data can then be used as additional training data for the translation system. Such methods generally have a very low yield leaving vast amounts of data that is only used for language modeling. These methods rely upon comparable corpora, that is, multiple corpora that are of the same gen- eral genre. In addition to this, documents can be comparable—two documents that are both on the same event or topic. Comparable documents occur because of the repetition of information across lan- guages, and in the case of news data, on the fact that stories reported in one language are often reported in another language. In cases where no direct trans- lation can be found for a source document, it …

Content truncated to 500 words