Improving Statistical MT through Morphological Analysis

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 676–683, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Improving Statistical MT through Morphological Analysis Sharon Goldwater Dept. of Cognitive and Linguistic Sciences Brown University sharon goldwater@brown.edu David McClosky Dept. of Computer Science Brown University dmcc@cs.brown.edu Abstract In statistical machine translation, estimat- ing word-to-word alignment probabilities for the translation model can be difficult due to the problem of sparse data: most words in a given corpus occur at most a handful of times. With a highly inflected language such as Czech, this problem can be particularly severe. In addition, much of the morphological variation seen in Czech words is not reflected in either the morphol- ogy or syntax of a language like English. In this work, we show that using morphologi- cal analysis to modify the Czech input can improve a Czech-English machine transla- tion system. We investigate several differ- ent methods of incorporating morphological information, and show that a system that combines these methods yields the best re- sults. Our final system achieves a BLEU score of .333, as compared to .270 for the baseline word-to-word system. 1 Introduction In a statistical machine translation task, the goal is to find the most probable translation of some foreign language text f into the desired language e. That is, the system seeks to maximize P(e|f). Rather than maximizing P(e|f) directly, the standard noisy chan- nel approach to translation uses Bayes inversion to split the problem into two separate parts: argmax e P(e|f) = argmax e P(e)P(f|e) (1) where P(e) is known as the language model and P(f|e) is known as the translation model. The limit- ing factor in machine translation is usually the qual- ity of the translation model, since the monolingual resources needed for training the language model are generally more available than the parallel corpora needed for training the translation model. Due to the difficulty in obtaining large parallel cor- pora, sparse data is a serious issue when estimating the parameters of the translation model. This prob- lem is compounded when one or both of the lan- guages involved is a highly inflected language. In this paper, we present a series of experiments suggesting that morphological analysis can be used to reduce data sparseness and increase similarity between lan- guages, thus improving the quality of machine trans- lation for highly inflected languages. Our work is on a language pair in which the input language (Czech) is highly inflected, and the output language (English) is not. We discuss in Section 5 how our methods might be generalized to pairs where both languages are highly inflected. The plan of this paper is as follows: In Section 2, we review previous work on using morphologi- cal analysis for statistical machine translation. In Section 3, we describe several methods for utilizing morphological information in a statistical translation model. Section 4 presents the results of our experi- ments using these methods. Sections 5 and 6 discuss the …

Content truncated to 500 words