Human Editing For Google’s Neural Machine Translation Of English News

The rapid development of computer technology has changed the translators’ working style and more and more translators choose to rely on human-computer interaction to improve the efficiency of translation.Google’s Neural Machine Translation(GNMT)adopts a neural network translation system and brings convenience and efficiency to translators in language conversion,which makes it the first choice of more and more translators.The translation of news report features accuracy and fluency.GNMT’s immediate service can assist and speed the procedures of translation but Google’s version,limited by its database or algorithm,is still full of different kinds of errors and cannot be used directly.So human editing is necessary to keep the translated version fluent and readable.This thesis uses Google’s neural network translation system as a tool,takes English to Chinese translation of news text as example and selects the bilingual news of the website “English Tips” of China Daily as the sample,and compares Google’s translation results with human translation of 216 original sentences of English news.The author marks various errors and establishes a three-grade error model based on lexical error,syntactic error and symbolic error for Google’s English-Chinese news translation.According to this model,the frequency and causes of those errors are summarized and analyzed.For the errors that can be predicted,corresponding pre-editing methods are proposed,such as pre-translating and ambiguity elimination,which largely reduce the time and amount of work of post-editing.For those errors that are not easy to be foreseen,post-editing methods,such as checking tense,checking terms and adding and omitting,are put forward in this paper to correct all the errors left,which prove to be effective in improving the quality and efficiency of news translation.This thesis analyzes the errors of GNMT on translation of English news and their causes.Effective methods of human editing are proposed based on the deficiencies,which provide new strategies for the efficiency improvement on human-computer interactive translation on English news.