Generating Complex Morphology for Machine Translation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 128–135, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Generating Complex Morphology for Machine Translation Einat Minkov∗ Language Technologies Institute Carnegie Mellon University Pittsburgh, PA, USA einatm@cs.cmu.edu Kristina Toutanova Microsoft Research Redmond, WA, USA kristout@microsoft.com Hisami Suzuki Microsoft Research Redmond, WA, USA hisamis@microsoft.com Abstract We present a novel method for predicting in- flected word forms for generating morpho- logically rich languages in machine trans- lation. We utilize a rich set of syntactic and morphological knowledge sources from both source and target sentences in a prob- abilistic model, and evaluate their contribu- tion in generating Russian and Arabic sen- tences. Our results show that the proposed model substantially outperforms the com- monly used baseline of a trigram target lan- guage model; in particular, the use of mor- phological and syntactic features leads to large gains in prediction accuracy. We also show that the proposed method is effective with a relatively small amount of data. 1 Introduction Machine Translation (MT) quality has improved substantially in recent years due to applying data intensive statistical techniques. However, state-of- the-art approaches are essentially lexical, consider- ing every surface word or phrase in both the source sentence and the corresponding translation as an in- dependent entity. A shortcoming of this word-based approach is that it is sensitive to data sparsity. This is an issue of importance as aligned corpora are an ex- pensive resource, which is not abundantly available for many language pairs. This is particularly prob- lematic for morphologically rich languages, where word stems are realized in many different surface forms, which exacerbates the sparsity problem. ∗ This research was conducted during the author’s intern- ship at Microsoft Research. In this paper, we explore an approach in which words are represented as a collection of morpholog- ical entities, and use this information to aid in MT for morphologically rich languages. Our goal is two- fold: first, to allow generalization over morphology to alleviate the data sparsity problem in morphology generation. Second, to model syntactic coherence in the form of morphological agreement in the target language to improve the generation of morphologi- cally rich languages. So far, this problem has been addressed in a very limited manner in MT, most typ- ically by using a target language model. In the framework suggested in this paper, we train a model that predicts the inflected forms of a se- quence of word stems in a target sentence, given the corresponding source sentence. We use word and word alignment information, as well as lexi- cal resources that provide morphological informa- tion about the words on both the source and target sides. Given a sentence pair, we also obtain syntactic analysis information for both the source and trans- lated sentences. We generate the inflected forms of words in the target sentence using all of the available information, using a log-linear model that learns the relevant mapping functions. As a case study, we focus on the …

Content truncated to 500 words