On The Quality Evaluation Of Five Machine Translations

Since the birth of machine translation(MT),the question whether MT can replace human beings has always existed,especially at present,MT is in a prosperous period.It prompts researchers to evaluate the quality of MT outputs.However,existing studies are somewhat subjective in the selection of research objects and text types,and also lack of theoretical guidance in the analysis of errors in MT outputs.With the development of network technology,the use of the Internet to publish questionnaires enables researchers to know more about the objective needs of MT users.In addition,linguistics is also developing.Under the guidance of Error Analysis,errors in MT outputs are classified and analyzed,which makes the evaluating results more scientific.This thesis takes five MTs as the research objects,among which Youdao,Google,Baidu and Tencent are frequently used by 96 users in the questionnaire.Deep L is a new and highly evaluated MT.There are four types of texts to be tested including literature,news,economy and technology,which are frequently translated by users.In the process of empirical research,a five-point scale is adopted to score 20 outputs with two evaluating criteria of intelligibility and fluency.Then,the quality of MT outputs is compared and evaluated in the form of tables.In the process of analyzing the MT outputs,it can be found that there are various errors.Under the guidance of Error Analysis Theory in linguistics,errors are analyzed from lexical,sentence and other aspects.At the same time,typical examples are selected for further analysis,and the sources of the errors are analyzed in the end.In addition,the starting point of the study on quality evaluation of MT outputs is the actual needs of users.Therefore,the way of making full use of MT is also given in the conclusion.It is found that in English-Chinese translation,Youdao performs the best in translating four types of texts.Deep L ranks the second,followed by Baidu.Tencent and Google perform the worst.MTs perform the best in translating news texts and the worst in translating literary texts.When translating economic texts and technological texts,the performance is average.There are still errors in the outputs,and errors at lexical and sentence aspects are more serious,among which context analysis and native expression are the difficulties for MT.However,the overall performance of MT is relatively optimistic.Finally,this thesis points out that MT can not be totally denied,and users can make full use of it by pre-editing and post-editing and using multiple MTs at the same time.