The Research On The Decodeing Algorithm In Statistical Machine Translation

During the last two decades, statistical machine translation has successfully outweighed other machine translation approaches in many evaluations, which makes it one of the hottest issues in the machine translation domain. Through a statistical analysis of the bilingual corpus, it is able to learn the translation knowledge automatically. Therefore, research on statistical machine translation is of great theoretical and practical significance. Decoding algorithm plays an important role to statistical machine translation. The aim of decoding is to find the result with the highest score according to a given statistical (translation) model in the space of target words. The performance of decoding algorithm has a direct impact on the quality and efficiency of translation. In this paper, studies are focused on the decoding algorithms of phrase and syntax-based statistical machine translation, respectively. According to the features of these models, efficient scoring strategies and heuristic functions are adopted and beam search algorithm is applied. Experimental results show that the presented approach achieves relatively high sores in evaluation and a satisfactory translation speed as well.This thesis is arranged as follows:(1) Research on the decoding algorithm of the phrase-based statistical translation model. Given the phrase-based translation model and the language model, consider the translation histories and the reasonable future estimations as the translation cost, using the beam search to cut down the lower scored hypothesis when they cover the same foreign words. It can cut down search scope and speedup the search to get the best translation.(2) Research on the decoding algorithm of the syntax-based statistical translation model. Based on the Yamada's syntax-based translation model and the language model, use a method similar with the parser to do the translation from the source language to the destination language. Beam search is also used in the decoding, keep the top-n production rules to do the next deduction, bottom-up to form a tree structure.(3) Experiments on the decoders. Evaluate on the 863 corpus, and...