Neural Machine Translation With Linguistic Information Integration

Along with the process of globalization,multinational economic and cultural exchanges have become increasingly frequent.In the wake of collision and integration of multi-cultures,how to break through language barriers,rendering machines accomplish automatic translation between various languages and finally achieving barrier-free communication has been peopled ambition for long time.With the new vigor of deep learning,the system of neural machine translation has made great progress in translation quality,and its compre-hensive performance has surpassed traditional statistical machine translation system signifi-cantly.Neural machine translation adopts an end-to-end encoder-decoder framework when modeling data,and does not combine abundant priori linguistic information.This paper attempts to integrate rich linguistic information into the encoder-decoder framework of neu-ral machine translation system to explore the content and methods of linguistic information integration.In proper order of word level,local level and sentence level,this paper integrates lin-guistic information into neural machine translation ascendingly to improve the performance of experimental system.In the process of integrating linguistic information into neural ma-chine translation model:Firstly,from the perspective of word level,the study integrates the part-of-speech information into the neural machine translation model.By adding part-of-speech sequences and sharing context vectors at the target,word sequences and part-of-speech sequences are jointly decoded in the decoding stage,and part-of-speech sequences then assist in the generation of target word sequences.Experimental results show that neural machine translation incorporating with part-of-speech information can improve the quality of translation output.Secondly,from the perspective of local level,the study integrates local relevance information into the neural machine translation model,and makes each word more associated with its context by adding neighbor relevance guidance at the source and the tar-get sides.Experimental results prove that:a)translation system integrated with neighbor relevance guidance respectively at source and target side is superior to the baseline;b)the system which merges neighbor relevance guidance from both source and target sides is also obviously better than the baseline.Lastly,on the basis of syntactic structure of sentences,the information of dependency tree and dependency relevance guidance are integrated into the neural system for the purpose of constructing long-distance syntactic relevance.Exper?imental results illustrate that the increase of dependency relevance guidance improves the quality of neural machine translation.In summary,the integration of linguistic information from word level to sentence level can effectively promote the performance of neural machine translation.