Research On The Quality Evaluation "Google Neural Machine Translation System"

Since the Industrial Revolution,human has rapidly developed into a prosperous scene.Now,mankind is facing another revolution,the intelligent revolution.The rapid rise of artificial intelligence(AI)has set off a new round of revolution,the AI era has came.As the core of cross-language communication,the demand for language translation services has increased dramatically with the development of world economic,trade and culture.Traditional translation is difficult for the needs of largescale translation.The development of information technology and computer hardware has promoted the development of machine translation.Many organizations are using machine translation(MT)to reduce translation costs.The automatic translation platform with neural network based on AI has made a huge leap in accuracy in the past two years.One of them is Google Translate.In September 2016,Google launched the new online translation platform GNMT(Google Neural Machine Translation System)to replace the previous Statistical Machine translation(SMT).GNMT is based on the end-to-end translation learning method.The accuracy of translation in languages such as English,French and German has reached 90%,and the accuracy of translation between Chinese and English is about 80%.In the past several years,Google Translate provides millions of translations across multiple language pairs every day.Nearly one billion people use it every month.However,the number of people who really understand its internal working is much less than before,and the quality of its translation is difficult to evaluate.This article discusses the application of GNMT in the translation of tourism texts from English to Chinese,discovers some main problems of machine translation in the translation of specific texts,and puts forward some suggestions for the translation system.This paper first briefly reviews the machine translation history and the emergence and evolution of GNMT,and evaluates the feasibility and practicality of using GNMT technology to translate English to Chinese in tourism texts;then it describes the types and development of machine translation evaluation,The new NMT system greatly surpasses the previous statistical machine translation in both fluency and readability.However,due to the problem of algorithm bias,excessive pursuit of fluency may reduce the faithfulness of the translation.GNMT optimizes the BLEU value method based on N-gram matching for statistical machine translation.The score may be high,but the translation quality is difficult to guarantee the same as the score.This requires the establishment of a new three-dimensional evaluation system.Then introduce and classify the tourism texts Lonely Planet Europe,which is the evaluation target of this thesis;on the framework of DQF dynamic evaluation theory,combined with the research of World Machine Translation(WMT)on machine translation quality,Establish a set of quick evaluation framework for machine translation based on DQF dynamic theory,first evaluate the overall readability of the mainstream translation system-Google Translation,and then analyze translation errors from three levels: language,terminology and accuracy to answer the following three questions:(1)As a general machine translation system,how good is the quality of GNMT for English-toChinese translation of particular text?(2)In the face of specific text,what are the typical mistakes that often occur in GNMT English-Chinese translation?(3)What are the GNMT improvement measures for these typical errors? The analysis result shows that the GNMT translation is 83.8 points,which is within the user’s acceptable range,in the case of the particular text and the target user.Typical errors mainly focus on the inaccurate translation of terms and failure to effectively deal with unregistered words;in terms of accuracy,the main errors are the problems of multiple translations,missed translations,and wrong translations;the main problems in language It is on the syntactic,sentence composition and collocation issues that the sentence does not conform to Chinese grammar.GNMT should further enhance the ability to recognize and understand natural language,adopt new algorithms and limiting mechanisms,and use some linguistic knowledge to optimize output to approach or reach the level of manual translation.