An Analysis Of Functional Equivalence In Google Neural Machine Translation From The Perspective Of Halliday’s Meta-function Theory

Since the birth of machine translation(MT),researchers have been in constant efforts to improve and optimize its quality;nowadays,various machine translation systems launched MT based on neural network,which makes the translated text more faithful in content and more fluent in language.Compared with human translation,the quality of Neural MT is still unsatisfactory,especially in terms of comprehensibility and accuracy.Machine translation,as an interdisciplinary study,includes linguistics,mathematics,and computational science.Among these disciplines,linguistics provides theoretical guidance to deal with language problems in MT outputs.Based on the role linguistics plays in MT study,this research,from the perspective of Halliday’s Meta-function theory,evaluates the functional equivalence between source text and translated text of Google Neural Machine Translation(GNMT).By analyzing the non-equivalence phenomenon,the author aims to explore the reasons for the translation non-equivalence.The research material is informative text regarding to environmental protection,selected for National Geographic,and the target of this research is the Chinese outputs of GNMT.Guided by Meta-function theory,the author compares the translated text with source text from ideational function,interpersonal function,and textual function,identifying the non-equivalence phenomena,and has a detailed analysis to finds out the reasons for these non-equivalence phenomena.The results show that in terms of the translation of informative text,(1)the ideational function has the highest non-equivalent translation proportion,accounting for70%,which severely affects readers’ understanding of the text.Specifically,among the six process,material process and relational process appear most frequently and also take the largest deviation proportion.For material process,deviation in the translation of Process has the highest probability,while for relational process,deviation in the translation of Participant occurs most frequently.(2)Non-equivalence in textual function takes a proportion of 21%,which mainly reflects in the long and complex sentence and inserted components,hindering readers’ comprehension of the passage to a great extent.(3)Non-equivalence phenomenon in interpersonal function is rare,at a rate of 9%,showing that compared with ideational and textual function,functional equivalence of GNMT in interpersonal function is well-achieved.There are several potential reasons for the above-mentioned non-equivalence phenomena.(1)The mechanization of GNMT results in literal translation,and the word-for-word translation makes machine unable to accurately retrieve the collocations,causing a deviation.(2)GNMT takes sentence as a translation unit,which indicated that the machine can only take the context of a sentence into consideration,lacking the contextual guidance of paragraphs or the whole passage.As a result,the translated text is inaccurate and rigid.(3)Machine is still weak in parsing long sentences,especially sentences with inserted clause or subordinate clause,which results in illogical and unintelligible translation.(4)Due to the limited corpora of GNMT as well as the uneven quality of data,the translation of terminology has a large deviation proportion,especially,the phenomenon that one same word having inconsistent translation appears with a high frequency.