Research On Introducing Linguistic Information In Neural Machine Translation

With the development of the globalization,international exchanges have become more frequent,which leads to that the barrier of language communication becomes more serious.Machine translation,which is a hotpot in the field of artificial intelligence and natural language processing,has become an important way to overcome the barrier and promote international and cross-cultural exchanges with efficient and low cost.Neural machine translation has achieved extensive attention due to the progress of deep learning.Compared with previous machine translation systems,neural machine translation systems only rely on parallel data-sets and does not use any linguistic infor-mation.In the training stage,input and output languages are mapped into two vector space which contain all information needed in translation process.However,paral-lel data-sets are manually annotated whose size is small and the parameters of neural networks are very large.This unbalances may lead to the hidden representations are not exact which will further cause some translation errors like under-translation,over-translation or semantic deviation.Linguistic information is a useful complement in this situation for guiding the translation process to generate more complete and accurate re-sults.This thesis focuses on the use of linguistic information in neural machine transla-tion on the aspects of the model structure and the external resource.The main works are as follows:1.In order to solve the problem that the output in neural machine translation is inac-curate.This thesis proposes a word prediction mechanism to constrain each state from the decoder by predicting all words should be translated in the future.By this way,all states can contain the future information of translation and the linguistic information contained in states becomes more complete.The accuracy of results from the proposed model is higher than the original model.2.Advanced neural machine translation does not model global sentence level repre-sentation explicitly which will lead to semantic or syntactic deviation.This thesis proposes to generate a global sentence level state in encoding state and fuse it into each state of the decoder dynamically.3.The linguistic representations only learned from parallel data-set may not good enough.This thesis proposes a novel structure to achieve linguistic information by large scale monolingual data.Then,two methods are proposed to utilize external linguistic representations to improve translation quality.4.On the other hand,in human-machine interaction situation,this thesis proposes a novel bi-directional interactive framework which can fully use the linguistic infor-mation provided from users,and further improve the translation quality.