Generating Case Markers in Machine Translation

Proceedings of NAACL HLT 2007, pages 49–56, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Generating Case Markers in Machine Translation Kristina Toutanova Hisami Suzuki Microsoft Research One Microsoft Way, Redmond WA 98052 USA {hisamis,kristout}@microsoft.com Abstract We study the use of rich syntax-based statistical models for generating gram- matical case for the purpose of machine translation from a language which does not indicate case explicitly (English) to a language with a rich system of surface case markers (Japanese). We propose an extension of n-best re-ranking as a method of integrating such models into a statistical MT system and show that this method substantially outperforms stan- dard n-best re-ranking. Our best perform- ing model achieves a statistically signifi- cant improvement over the baseline MT system according to the BLEU metric. Human evaluation also confirms the re- sults. 1 Introduction Generation of grammatical elements such as in- flectional endings and case markers is an impor- tant component technology for machine transla- tion (MT). Statistical machine translation (SMT) systems, however, have not yet successfully in- corporated components that generate grammati- cal elements in the target language. Most state- of-the-art SMT systems treat grammatical ele- ments in exactly the same way as content words, and rely on general-purpose phrasal translations and target language models to generate these ele- ments (e.g., Och and Ney, 2002; Koehn et al., 2003; Quirk et al., 2005; Chiang, 2005; Galley et al., 2006). However, since these grammatical elements in the target language often correspond to long-range dependencies and/or do not have any words corresponding in the source, they may be difficult to model, and the output of an SMT system is often ungrammatical. For example, Figure 1 shows an output from our baseline English-to-Japanese SMT system on a sentence from a computer domain. The SMT system, trained on this domain, produces a natu- ral lexical translation for the English word patch as correction program, and translates replace into passive voice, which is more appropriate in Japanese. 1 However, there is a problem in the case marker assignment: the accusative marker wo, which was output by the SMT system, is completely inappropriate when the main verb is passive. This type of mistake in case marker as- signment is by no means isolated in our SMT system: a manual analysis showed that 16 out of 100 translations had mistakes solely in the as- signment of case markers. A better model of case assignment could therefore improve the quality of an SMT system significantly. S: The patch replaces the .dll file. O: 修正プログラムを.dll ファイルが置き換えられます。 shuusei puroguramu-wo .dll fairu-ga okikae-raremasu correction program-ACC dll file-NOM replace-PASS C: 修正プログラムで.dll ファイルが置き換えられます。 shuusei puroguramu-de .dll fairu-ga okikae-raremasu correction program-with dll file-NOM replace-PASS Figure 1: Example of SMT (S: source; O: output of MT; C: correct translation) In this paper, we explore the use of a statisti- cal model for case marker generation in English- to-Japanese SMT. Though we focus on the gen- eration of case markers in this paper, there are many other surface grammatical phenomena …

Content truncated to 500 words