Training Data Modification for SMT Considering Groups of Synonymous Sentences

摘要

Generally speaking, statistical machinetranslation systems would be able to attainbetter performance with more training sets.Unfortunately, well-organized training setsare rarely available in the real world. Consequently,it is necessary to focus on modifyingthe training set to obtain highaccuracy for an SMT system. If the SMTsystem trained the translation model, thetranslation pair would have a low probabilitywhen there are many variations for targetsentences from a single source sentence.If we decreased the number of variationsfor the translation pair, we could constructa superior translation model. This paper describesthe effects of modification on thetraining corpus when consideration is givento synonymous sentence groups. We attemptthree types of modification: compressionof the training set, replacement ofsource and target sentences with a selectedsentence from the synonymous sentencegroup, and replacement of the sentence ononly one side with the selected sentencefrom the synonymous sentence group. As aresult, we achieve improved performancewith the replacement of source-side sentences.