| Paraphrases are the words or phrases which are alternative ways of conveying the same meaning.Paraphrases in different domains have a wide range of applications in natural language processing.In automatic machine translation evaluation,the domain of paraphrase has a great influence on the performance of automatic evaluation system.When the domain of paraphrase is similar to that of the reference,the matching accuracy between the machine translation and the reference translation will be increased,thus the performance of the automatic evaluation will be improved.Conversely,the performance will be lower.The loss of corpus and paraphrase match deviation are likely to happen in machine translation evaluation using paraphrase.To deal with this problem,this paper proposes a method using specific-domain paraphrae related to the reference to improve the automatic evaluation performance.First,we cluster general-domain monolingual corpus into general-domian documents,then,filter into the specifc-domain sub-corpus via improve M-L approach.Finally,we extract the specific-domain paraphrase table from the sub-corpus by Markov network model.The extracted paraphrase table is applied to automatic evaluation metrics to improve word match.In order to validate the proposed method,we conduct extensive experiments on WMT'14,WMT'15 and WMT'16 tasks,and the results show that the specific-domain paraphrase method is not only simple in construction with less language restrictions,but also effective in increasing the matching accuracy between the machine translation and reference translation.The further experimental analysis of the method indicates that using specific-domain paraphrase greatly improves the matching precision of synonyms,and ultimately improves the performance of automatic machine translation evaluation method.It is encouraging that we ranked the first place by two indicators using the proposed method in the WMT'16 Metrics task according to the official results. |
PDF Full Text Request