Flexible Statistical Construction of Bilingual Dictionaries

Most previous systems for constructing a bilingual dictionary from a parallel corpus have depended on an iterative algorithm, using word translation probabilities to align words in the corpus, and using word alignments to estimate word translation probabilities, and repeating until convergence. While this approach produces reasonable results, it is computationally slow, limiting the size of the corpus that can be analysed and the size of the dictionary produced. We propose a non-iterative approach for producing a uni-directional bilingual dictionary which, while less accurate than iterative approaches, is far quicker, allowing larger corpora to be processed in reasonable time. The approach also allows on-the-fly estimation of translation likelihoods between a pair of terms, meaning that a translation dictionary can be generated with the n most frequent terms in an initial pass, and the translation likelihood of infrequent terms can be calculated as encountered in real documents.