Improved Word Alignment Using a Symmetric Lexicon Model

Improved Word Alignment Using a Symmetric Lexicon Model Richard Zens and Evgeny Matusov and Hermann Ney Lehrstuhl f¨ur Informatik VI, Computer Science Department RWTH Aachen University D-52056 Aachen, Germany {zens,matusov,ney}@cs.rwth-aachen.de Abstract Word-aligned bilingual corpora are an important knowledge source for many tasks in natural language processing. We improve the well-known IBM alignment models, as well as the Hidden-Markov alignment model using a symmetric lex- icon model. This symmetrization takes not only the standard translation direc- tion from source to target into account, but also the inverse translation direction from target to source. We present a the- oretically sound derivation of these tech- niques. In addition to the symmetriza- tion, we introduce a smoothed lexicon model. The standard lexicon model is based on full-form words only. We propose a lexicon smoothing method that takes the word base forms explicitly into ac- count. Therefore, it is especially useful for highly inflected languages such as Ger- man. We evaluate these methods on the German–English Verbmobil task and the French–English Canadian Hansards task. We show statistically significant improve- ments of the alignment quality compared to the best system reported so far. For the Canadian Hansards task, we achieve an improvement of more than 30% rela- tive. 1 Introduction Word-aligned bilingual corpora are an impor- tant knowledge source for many tasks in nat- ural language processing. Obvious applica- tions are the extraction of bilingual word or phrase lexica (Melamed, 2000; Och and Ney, 2000). These applications depend heavily on the quality of the word alignment (Och and Ney, 2000). Word alignment models were first introduced in statistical machine translation (Brown et al., 1993). The alignment describes the mapping from source sentence words to target sentence words. Using the IBM translation models IBM-1 to IBM-5 (Brown et al., 1993), as well as the Hidden-Markov alignment model (Vogel et al., 1996), we can produce alignments of good quality. In (Och and Ney, 2003), it is shown that the statistical approach performs very well compared to alternative approaches, e.g. based on the Dice coefficient or the com- petitive linking algorithm (Melamed, 2000). A central component of the statistical trans- lation models is the lexicon. It models the word translation probabilities. The standard training procedure of the statistical models uses the EM algorithm. Typically, the models are trained for one translation direction only. Here, we will perform a simultaneous training of both translation directions, source-to-target and target-to-source. After each iteration of the EM algorithm, we combine the two lexica to a symmetric lexicon. This symmetric lex- icon is then used in the next iteration of the EM algorithm for both translation directions. We will propose and justify linear and loglin- ear interpolation methods. Statistical methods often suffer from the data sparseness problem. In our case, many words in the bilingual sentence-aligned texts are singletons, i.e. they occur only once. This is especially true for the highly inflected lan- guages such as German. It is hard to obtain reliable estimations of the translation proba- bilities for these rarely occurring words. …

Content truncated to 500 words