Research On Automatical Evaluation Technology Of Translation Quality In Chinese-English Translation

The present automatic evaluation methods of translation generally aim at machine translation. The comparison of the performance of different systems is achieved by calculating the similarity of candidate translations and reference translations, but the strategy is infeasible for human translation or machine-assisted translation without reference translations. However, great economic and time cost can be consumed by adopting human evaluation. Thus, how to achieve an effective automatic evaluation method becomes a research hotspot for human translation or machine-assisted translation.In this thesis, a qualitative evaluation method without reference translation based on SVM is proposed by researching the present evaluation methods for machine translation. Different from the traditional methods of calculating the similarity of candidate translations and the references for evaluation, we first classify the sentences of the translation according to the existence of errors and error types, then give scores according to the classification result and carry out qualitative evaluation. Experiments indicate that the method is relevant to human evaluation to some degree.The work of this thesis also includes:First, a solution strategy of article choice problems based on conditional random fields is proposed. Aiming at the problem that traditional methods based on rules or machine learning only consider the selection of articles before noun phrases, we describe the problem as a sequence labeling task in formalization, fully utilize the multilayer characteristics of words, part of speech, etc., and introduce mutual information of adjacent words to achieve the selection of articles by the conditional random fields.Second, word misspelling check based on dictionaries is realized, discovered word mistakes can be prompted by adopting the thought based on edit distance, and large quantities of translations can be automatically checked.Third, on the basis of analyzing the constructional characteristic of patent translation sentences, the check of subject-predicate consistency is achieved. The central word of the subject part corresponding to the predication verb can be analyzed by constructing the syntactic tree of the sentences so as to judge the subject-predicate consistency error.The work above is applied to the system for translation sampling inspection and checking after translation in translation work, which indicates that mistakes in the translation can be discovered effectively and the quality testing efficiency of translation can be improved to a certain degree.