Over the past decade, machine translation has been greatly developed. A variety of different para-digms for statistical machine translation (SMT) have been proposed, including phrase-based SMT model, syntax-based SMT model etc. Each model has its strength and weakness. It would be a very meaningful work to integrate the advantages of multiple translation engines and overcome their shortcomings. The combination of machine translation system extracts useful information from the outputs of multiple machine translation engines to get the final consensus translation. It has been widespread concern as an effective way to improve the quality of machine translation. Thus, the research on machine translation system combination has important theoretical and practical value. Under the framework of word-level system combination, and taking the Chinese-to-English and English-to-Chinese machine translation systems as the experimental platform, we deeply study the methods of system combination and put it into practice. The major contributions are listed as follows: 1. We present a new approach to word reordering alignment and put it into practice The alignment between paired monolingual sentences is an important process for word-level system combination. We present a word reordering alignment (WRA) approach for combination of SMT systems in this paper. Different from the previous approaches based on edit distance, such as WER or TER, our WRA approach directly shifts the word sequences of the translation hypothesis to the correct location within the translation hypothesis. In our approach, the continuous word sequences are first found and replaced by some variables. Then we align the variables and words identical to each other in the two sentences and detect the cross alignment that should be reordered. According to the cross alignment, the detected word sequences are shifted to the correct position and dynamic programming are exploited to align the sentences after reordering. The experiments on newswire translation domain and spoken language translation domain show that the approach can significantly improve the translation quality. 2. We compare the diffent approaches to system combination There are three different levels of systems combination methods for machine translation, including sentence-level system combination, phrase-level system combination, and word-level system combi-nation. To compare different system combination approaches, we conduct experiment on s...
修改评论