A Linguistic Approach To Error Types Of Google Neural Machine Translation

Neural machine translation(NMT)is an end-to-end structure with the encoder and the decoder and it comprehends mapping between natural languages through deep lear ning of neural network.NMT,as the latest stage of machine translation,distinguishes itself from traditional phrase-based machine translation for its significant improvement on fidelity and fluency of its outputs.Google Neural Machine Translation(GNMT)has made several technological breakthroughs.Nevertheless,many problems remain to be solved to improve quality of GNMT outputs.The present study is conducted from the perspective of linguistics,using the GNMT output of a text extracted from An Excerpt from Investment-Related Policies of Guizhou Province,a government document.The source text is translated as a whole on Google Translate(https://translate.google.cn/).Under the framework of the error analysis theory and based on register analysis of the source text,comparison of ideational meaning and interpersonal meaning at clause level as well as comparison of textual meaning at text level is made between the source text and the GNMT output,so as to identify errors in ideational,interpersonal and textual meanings in the output.With reference to the integrated error categories adopted by five institutions/companies,errors identified in the present study are categorized from the aspects of fidelity,fluency and style into three categories and twent y subcategories.When causes of errors are explored,there are three major findings.Firstly,Chinese and English,which belong to two language families,distinguish themselves from each other with the characteristics of parataxis and hypotaxis respectively.In Chinese phrases and clauses are usually placed one after another without conjunctions to show their relationships,while in English components within a sentence are connected by prepositions,conjunctions,subordinate clauses,etc.;lexical and grammatical devices such as conjunctions,references,substitutions,ellipses,synonyms/antonyms,repetition and the like are widely used to connect sentences and to show their logic relationships.Such prominent difference between Chinese and English poses difficulties for GNMT to produce effective text translation.Secondly,although being much more advanced than traditional machine translation,GNMT still deals with language that is highly formalized and abstracted;language processing in GNMT still focuses on sentence structures as well as generation and transformation of sentences,giving little play to contexts among sentences and in the whole text.Therefore,GNMT is weak in utilizing contexts for correct recognition of the source message,reasonable sentence division,accurate diction and appropriate style.Besides,the unstable performance of GNMT also leads to errors like inconsistent translation of the same expressions,and unreasonable capitalization of words within sentences.Thirdly,not all the translation options in the training corpus are good in quality,and sometimes there is more than one translation option for a particular phrase/expression,which also leads to the unsatisfactory performance of GNMT.Based on the above findings,suggestions for improving quality of GNMT outputs are proposed from perspectives of human intervention and technological advance of machine translation,including pre-editing,post-editing as well as technological advance.Effective pre-editing include subject adding and sentence reorganization to reveal logic relationships among components;post-editing is the necessary process for correcting errors and polishing machine outputs;concerning technological advance,suggestions include unifying translations for widely used terms and setting rules for translating sentences in fixed patterns.Moreover,human translators are suggested to work rather than compete with machine translation systems and to play a more creative role in their interaction with machine translation.