Extraction And Matching Of Translation Template In EBMT System

EBMT (Example-Based Machine Translation) systems are based on large scale example corpus in traditional, having the defect of low precision of matching. Translation template can solve the problem of data sparsity, large storage space and low matching precision of examples. The research in this paper focuses on the automatic Translation template extraction and matching based on the example corpus.The translationtemplate defined in this paper is based on the result of the shallow parsing, including the main verb identification, prepositional phrase identification and chunk parsing. Shallow parsing can recognize more reliable result than full parsing, and makes full syntax parsing easier.Extraction and matching of templates are the most important problems of Template-Based Machine Translation. The extraction module extracts the sentence frame, prepositional phrase and chunk templates from the result of shallow parsing. The templates are storaged independtly and linked by keyword-indexing in database. The matching module searches the most similar template for input sentence in database, with the information of syntactic structure and lexical meaning of the sentence. The templates matching algorithm gets the searching result by using key word as the static threshold, distance and similarity score as the dynamic threshold.The close test on sentence level and open test on chunk level based on the templates database builded on 2386 sentences show promising results: the precisions are above 94.98% and 94.85%. The results indicate that it's feasible to use the translation template applied in this paper in EBMT systems.An EBMT translation engine for NiHao Chinese-Japanese translation system is applied in this paper. The template definition and working flow of the engine are detailedly designed and the preparatory experiment has got a good testing result.