Study On Chinese-english Neural Machine Translation Incorporating Discourse Knowledge

Machine translation is the key technology for human beings to break the language barrier and achieve language interoperability.It is also one of the most important research directions in the field of natural language processing and artificial intelligence.Therefore,machine translation research not only has important application value,but also has important theoretical significance.The neural machine translation has made great progress in capturing the semantic association between words and sentence translation,and has become the mainstream machine translation technology.Many achievements have emerged in the research of incorporating discourse context information in neural machine translation.However,the current document-level neural machine translation has shortcomings in modeling the discourse structure relationship between sentences or clauses in a discourse,resulting in the translation model often generating discourse translations with poor coherence and readability.Although researchers have proposed many methods to solve this problem by using context information,most of these methods rely on the reference of large-scale serialized discourse context information to obtain discourse knowledge.The modeling lacks pertinence,and the way of incorporating discourse knowledge is relatively shallow,which can not effectively obtain the deep semantic knowledge of discourse structure.The methods are less effective in incorporating discourse structure knowledge,especially in facing Chinese,a paratactic language with more implicit discourse relations.Therefore,it is of great significance to explore a more accurate and effective discourse structure knowledge modeling and fusion mechanism in neural machine translation to improve the coherence of translation.This thesis focuses on the central problem of how to model and incorporate discourse structure knowledge in Chinese-English document-level neural machine translation,and proposes a series of solutions by combining the discourse analysis theory.The research methods and results of the thesis are summarized as follows.First,this thesis proposes a Chinese-English neural machine translation method incorporating clause alignment knowledge.At present,the problem of modeling discourse structure knowledge in neural machine translation is that there is no clear basic modeling unit,the modeling lacks pertinence,and the way of integrating discourse structure knowledge is relatively shallow.To address this problem,this thesis proposes to take clause as the basic modeling unit of discourse structure,and deeply study the Chinese-English bilingual discourse structure alignment based on clauses,as well as the method of incorporating the clause aligned discourse structure knowledge into the translation model.on the one hand,this thesis constructs a large-scale Chinese-English discourse structure parallel corpus based on clause alignment through using a combination of manual and automatic annotation,which provides explicit discourse structure alignment knowledge for neural machine translation model;On the other hand,this thesis designs a Chinese-English neural machine translation model incorporating clause alignment knowledge.The model effectively incorporates clause alignment knowledge by enhancing the sentence semantic representation based on clause components at the source side as well as enhancing clause alignment learning at the source and target sides.The experimental results show that the proposed method can effectively model and integrate the knowledge of discourse structure,and improve the performance of the document-level neural machine translation model.Second,this thesis proposes two Chinese-English neural machine translation methods incorporating the knowledge of discourse hypotactic structure alignment.At present,document-level neural machine translation is lack of modeling and fusion of discourse master-slave structure knowledge at the semantic level,which leads to the frequent errors of hypotactic structure transformation between the source language and the target language when the model is faced with long and complex sentences,which seriously affects the semantic coherence of discourse translation.This problem is particularly prominent in Chinese-English discourse machine translation due to the differences between the Chinese and English language forms.To address this problem,this thesis explores the modeling of discourse master-slave structure knowledge in Chinese-English neural machine translation,and proposes two neural machine translation methods that incorporate the knowledge of discourse hypotactic structure alignment.Specifically,this thesis first constructs a large-scale parallel corpus of Chinese and English discourse master-slave structures through using a combination of manual and automatic annotation,which provides rich discourse master-slave structure alignment knowledge for neural machine translation.On this basis,this thesis further proposes a neural machine translation method incorporating discourse hypotactic structure coding and a multi-encoder neural machine translation method incorporating discourse hypotactic structure alignment knowledge,and explores the method of incorporating discourse hypotactic structure alignment knowledge into neural machine translation from two aspects of coding strategy and model structure respectively.The experimental results show that the method proposed in this thesis can effectively model and incorporate the discourse hypotactic structure alignment knowledge,and then improve the performance of neural machine translation model.Third,this thesis proposes a document-level neural machine translation method based on multi-granularity fusion and multi-task learning.One major challenge faced by document-level neural machine translation is the scarcity of document-level training data.Thus,how to train a neural machine translation model to learn the semantic knowledge of discourse structure needs to be solved in the current machine translation research.To address this problem,this thesis proposes a Chinese-English document-level neural machine translation method based on multi-granularity fusion and multi-task learning.On the one hand,three levels of granularity for discourse context structural information,including clause,sentence and text,are introduced into neural machine translation to learn discourse structure knowledge from the perspective of multi-granularity fusion.On the other hand,two auxiliary task of discourse coherence modeling,sentence coherence detection and sentence restoration,are added to the translation task,so that the translation task and the auxiliary task are jointly trained in the way of parameter sharing,and the auxiliary task guides the translation model to fully learn additional discourse structure translation knowledge.The combination of the two aspects further enhances the performance of neural machine translation model.The experimental results show that the proposed method can help the model fully learn the knowledge of discourse structure under the condition of limited data resources,and then improve the translation effect of discourse translation.To sum up,focusing on the deficiencies of discourse structure knowledge modeling and fusion in document-level neural machine translation,this thesis proposes a series of solutions in clause alignment knowledge modeling and fusion,discourse hypotactic structure alignment knowledge modeling and fusion,and multi-granularity fusion and multi-task learning.Meanwhile,the discourse structure parallel corpora constructed in this thesis fill the gap of discourse structure data resources for machine translation.The experimental results show that the proposed methods can effectively improve the coherence of discourse translation and the translation quality of the model.