Improving Word Alignment Quality using Morpho-syntactic Information

Improving Word Alignment Quality using Morpho-syntactic Information Maja Popovi´c and Hermann Ney Lehrstuhl f¨ur Informatik VI - Computer Science Department RWTH Aachen University Ahornstrasse 55 52056 Aachen Germany {popovic,ney}@cs.rwth-aachen.de Abstract In this paper, we present an approach to include morpho-syntactic dependencies into the training of the statistical alignment models. Existing statistical translation sys- tems usually treat different derivations of the same base form as they were indepen- dent of each other. We propose a method which explicitly takes into account such in- terdependencies during the EM training of the statistical alignment models. The eval- uation is done by comparing the obtained Viterbi alignments with a manually anno- tated reference alignment. The improve- ments of the alignment quality compared to the, to our knowledge, best system are reported on the German-English Verbmobil corpus. 1 Introduction In statistical machine translation, a translation model Pr(fJ 1 |eI 1) describes the correspondences between the words in the source language sen- tence fJ 1 and the words in the target language sentence eI 1. Statistical alignment models are created by introducing a hidden variable aJ representing a mapping from the source word1 fj into the target word eaj. So far, most of the statistical machine translation systems are based on the single-word alignment models as described in (Brown et al., 1993) as well as the Hidden Markov alignment model (Vogel et al., 1996). The lexicon models used in these systems typically do not include any linguistic or con- textual information which often results in inad- equate alignments between the sentence pairs. In this work, we propose an approach to im- prove the quality of the statistical alignments by taking into account the interdependencies of different derivations of the words. We are get- ting use of the hierarchical representation of the statistical lexicon model as proposed in (Nießen and Ney, 2001) for the conventional EM training procedure. Experimental results are reported for the German-English Verbmobil corpus and the evaluation is done by comparing the ob- tained Viterbi alignments after the training of conventional models and models which are using morpho-syntactic information with a manually annotated reference alignment. 2 Related Work The popular IBM models for statistical ma- chine translation are described in (Brown et al., 1993) and the HMM-based alignment model was introduced in (Vogel et al., 1996). A good overview of all these models is given in (Och and Ney, 2003) where the model IBM-6 is also introduced as the log-linear interpolation of the other models. Context dependencies have been introduced into the training of alignments in (Varea et al., 2002), but they do not take any linguistic infor- mation into account. Some recent publications have proposed the use of morpho-syntactic knowledge for statisti- cal machine translation, but mostly only for the preprocessing step whereas training procedure of the statistical models remains the same (e.g. (Nießen and Ney, 2001a)). Incorporation of the morpho-syntactic knowl- egde into statistical models has been dealt in (Nießen and Ney, 2001): hierarchical lexi- con models containing base forms and set …

Content truncated to 500 words