Error Analysis Of English-chinese Machine Translation Output Under The Framework Of Clause Complex Theory

As deep learning was introduced into machine learning,neural machine translation(NMT)approach based on deep learning was successfully put into use in machine translation(MT)systems.In 2016,with a rapid development,neural machine translation took the place of statistical machine translation and became the kernel technology of commercialized MT systems,such as Google Translate.The development of machine translation promotes the advances of machine translation evaluation which pushes forward machine translation’s growth on the other way around.Human evaluation can not only judge the generation of machine translation,but also detect linguistic problems beneath the words and phrases when automatic evaluation primarily produces a score.Manual evaluation helps looking for a breakthrough for the improvements of machine translation quality.At present,there are a few studies on human evaluation of NMT based English-Chinese machine translation output.And according to related works reviewed in Chapter two,machine translation evaluation has focused on word,phrase,syntax and discourse,the assessment of machine translation has not yet extended to the level of clause complex.As a basic unit of language analysis,clauses,longer than phrases and shorter than sentences,constitute clause complexes in certain patterns.As the Clause Complex Theory based on NT clause was constructed,this research tries to observe the difficulties of English-Chinese machine translation on the clause complex level in particular.The study aims to uncover the categories of clause-complex level errors of English-Chinese machine translation output as well as their distribution due to the differences of English and Chinese clause complexes under the guidance of clausecomplex theory.Therefore,the author puts forward three specific research questions: 1.What kinds of clause-complex level errors could be perceived in the output of English-Chinese MT systems,such as Google Translate,Niutrans,Sougou Translate and Baidu Translate? 2.How do the subcategories of clause-complex level errors distribute in statistics? 3.What are the differences between English and Chinese clause complex structures that are related to the clause-complex level errors observed in the chosen MT systems’ output?The study adopts quantitative and qualitative methods.The author randomly selected 497 sentences from 20 English texts from Penn Treebank and input them into four mainstream online machine translation systems,such as Google Translate,Baidu Translate,Sougou Translate and Niutrans for collecting Chinese translations,which was completed from January 2017 to July 2017.The original and the generated sentences were parsed into NT clauses and presented in a Newline-Indent Schema.And next clause-complex level errors were annotated and further classified.Analysis with regards to the errors was carried out so as to reveal the problems on clause complex level.The study finds out that the clause-complex level errors mainly include four categories,such as context-sensitive word errors,naming-telling relation errors,logical relation errors and quotation errors.According to statistics,almost one clause-complex level error might appear in average two English sentences.This phenomenon has something to do with the structural differences between English and Chinese clause complex in terms of naming sharing relation and logic semantic relation.The study reveals the error types and their distribution on clause-complex level and analyzes the possible causes from the perspective of structural differences between English and Chinese clause complex.The findings of the study not only serve as a reference for improving machine translation quality but also contribute to the error analysis studies.