Machine Translation Based On Phrase Template

Generalization example in EBMT is an improvement to tranditional EBMT. It uses translation templates to replace examples, which not only improves the matching rate of the template to some extent, but also keeps high quality of translation. It's proved to be an efficient translation method.A strategy based on phrase template is proposed in this paper, which devides the template into two levels: phrase template and frame template. The research work also includes syntactic parsring, and it's introduced to extract the noun phrases from the sentences. Besides, a model of similarity computing method is proposed.The work contains two phases: training and translating. In the training phase, Chinese and Japanese noun phrases are extracted from a parallel bilingual corpus with syntactic parser, which devides the corpus into two parts: frame corpus and phrase corpus. By using the concept dictionary, we build the two corpuses into the phrase template library and frame template library. And in the translating phase, the target sentences are generated by combining phrase template and frame template: frame template to translate the frame of the input sentence and phrase template for the complicated noun phrases in the sentence. The similarity algorithm is based on the sequence method and it's trained with NIST to get a better parameter.Experimental result shows that this method obtains higher matching rate than the method of generalization EBMT while only requires a smaller quantity of templates. Its matching rate is 77.0% higher than the generalization's 51.5%.The approach described in this paper has brought a new idea to the research of EBMT, and it can also be applied in the machine translation system and other field.