A Cognitive Construal Perspective On Comparison Of Linguistic Features In Machine Translation And Human Translation

Machine translating software,as one of the most popular artificial intelligence products,is the application of translation and artificial intelligence.At present,the recursive neural network technology combines deep learning with translating process,modeling in terms of sentences,and through encoding and decoding to deconstruct the original text and reconstruct the translation.Theoretically,modeling in terms of sentence could construe the contextual meaning or collocative meaning,which could improve the translation quality efficiently.However,in humanity text translation,without precise cognition in contextual meaning,the accuracy is inferior to that in informative text.This research bases on the construal theory proposed by Langacker,focusing on the differences in construal dimensions between machine and human translation.The research questions are as followed:(1)Do machine translations differ from human translations in construal dimensions? If yes,what are the differences?(2)What is the distribution difference in terms of construal dimensions between two kinds of translations? And explore the possible causes for these differences,attempting to provide optimization suggestion for machine translating system in the future.The original text is a Chinese-English translation task,a part of the final examination for non-English major freshman in China.Human translation materials are students' answers.Machine translation materials are directly generated by Google Neural Machine Translation,Youdao Neural Machine Translation and Baidu Neural Machine Translation.This research adopts quantitative analysis and qualitative analysis simultaneously.Quantitative analysis is adopted in frequency statistics of each expression in human translation for the same original text.The most frequent three correspondences could represent the universality of human translation.Qualitative analysis is adopted in case study,exploring the differences between machine translation and human translation in construal dimensions.There are 25 comparative cases reflecting 31 differences in all,involving 14 differences in prominence dimension,6 in perspective dimension,8 in specificity dimension and 3 in background and scope dimension.Through comparative analysis,it is found that:(1)In prominence dimension,machine translating systems model in terms of sentence but segment sentence components in terms of word or phrase,which could cause language errors in pragmatics;(2)In perspective dimension,machine translating systems have set perfective marker,but lack parallel structure marker,while human translators could construe the parallel structure marker,but have various cognitions for “?” [liao],a general perfective marker;(3)In specificity dimension,machine translating system could operate reduction to avoid repetition according to output translation.However,in most occasion,machine translating systems stick to corresponding principle,avoiding reduction or amplification as far as possible;(4)In background and scope dimension,machine translating systems prefer neutral lexicons,avoiding breaking original background and scope constraints.Based on the cases analysis and discussion,this paper puts forward some suggestions for the optimization of machine translating system from the perspective of theoretical linguistics:(1)On the basis of sentence modeling,machine translating system could encode and decode sentence components in terms of chunks rather than single word or phrase,reconstructing the collocative meaning or contextual meaning precisely;(2)Add parallel structure marker in encoding and decoding process to improve the discourse cohesion;(3)Set up bilingual or multilingual parallel corpus for national cultural concepts,providing lingual materials for amplification in machine translations;(4)Mark appraisal lexicons in corpora.