Automatic Acquisition of Class-based Rules for Word Alignment

Automatic Acquisition of Class-based Rules for Word Alignment Sur-Jin Ker and Jason J. S. Chang Department of Computer Science National Tsing Hua University Hsinchu, Taiwan Abstract In this paper, we describe an algorithm for aligning words with their translation in a bilingual corpus. Existing algorithms require enormous bilingual data to train statistical word-to-word translation models. Using word-based approach, frequent words with consistent translation can be aligned at a high precision rate. However, less frequent words or words with diverse translations usually do not have statistically significant evidence for confident alignment. Incomplete or incorrect alignments consequently result. Our algorithm attempts to handle the problem using a hierarchical class-based approximation of translation probabilities. The translation probabilities are estimated using class-based models on 3 levels of specificity. We found that the algorithm can provide translation probability for more word pairs at the cost of slightly lower degree of precision, even when a small corpus was used in training. We have achieved an application rate of 81.8% and precision rate of 93.3%. The algorithm also offer the advantage of producing word-sense disambiguation information. 1. Introduction Much of the recent interests in bilingual corpora was initiated by Brown et al. (1990). They advocated a new approach to machine translation in which the bilingual corpora are aligned to reveal the mapping between text in one language and its translation in another language. This mapping is formally represented as statistical machine translation model. This model can be understood as a word by word model for generating (language model) of simple sentences S and translation (translation model) S to a sentence T in another language. The i-th word(s) s in S is considered to connect with its translation, the j-th word(s) t in T. Under the model, the probability of the connection (s, t) can be obtained by considering three aspects: lexical translation (t(t I s), the relation of s producing t in translation), fertility O('d*, the relation of change in number of words), and distortion (dis(i I j,1, m), the relation between the position i,j of s, t and the respectively length 1, m of S, 7). Probabilities are associated with these three relations of connection. A bilingual corpus annotated with these connection information can be utilized to estimate the parameters in a statistical translation model. Subsequently, the model can be used together with a bigram or trigram language model in machine translation. We present a rule-based algorithm for word alignment. We refer to this algorithm as SenseAlign. It relies on an automatic procedure for the acquisition of class-based rules. It does not employ word by word translation probabilities to identify alignment; nor does it - use a lengthy iterative EM algorithm for finding such probabilities. The algorithm attempts to handle the problem of undersampling by approximating word-to-word translation probability using hierarchical classifications of words. The translation probabilities are estimated using class-based rules on different levels of specificity. We found that the algorithm can provide translation probability for more source-target word pairs at the cost of slightly …

Content truncated to 500 words