The Use Of Multi-layer Alignment Framework For Instance-based Machine Translation Research

Machine Translation is one of the most important subjects in Natural Language Processing, with the development of Internet, more and more people want to find a way to communicate with each other. National Institute of Standards and Technology of US promotes the competition of machine translation from 2002. There are also many professional conferences in different countries including China. A lot of research center in big companies and famous universities join the competition and improve the quality of machine translation. But there still exist some problems in this field, for example, phrase alignment, re-order words, automatic evaluation of Machine Translation. Under the background, we do this research on Machine Translation.Our contribution is proposing a novel alignment framework, which is named Multi-layer alignment framework and can apply to Example-based Machine Translation (EBMT) easier. It contains three different types of alignments, syntax information and relatedness parameter. We implemented this framework, and finished a prototype EBMT System according to the framework. We also considered the expansibility and completeness of the Multi-layer alignment framework, and give some alternative schemes, definitions and interfaces for it.The experiments indicate that our framework got a good quality for alignment, especially in alignment intensity and relatedness parameter, we got 90% accuracy. When we used our EBMT system to do NIST test, we got 0.2966 in BLEU(BiLingual Evaluation Understudy) evaluation. It proved that multi-layer alignment is good for translation.