A Study On Discourse Parsing And Application On Neural Machine Translation

An end-to-end discourse parser is a system to distinguish discourse relations in free text,including explicit relations and implicit relations.The differece between explicit and implicit ones lay in wheather the connectives appear in the text explicitly.Discourse connectives play an important role in identifying the discourse relation between two textual units as well as locating these units,then identifying the relation between sentences or even paragraphs.A complete discourse parser contains several parts,including identifying connectives,identifying the relation between textual units,identifying the textual spans which connected by connectives.But no matter about the explicit nor implicit ones,the result of each part is not satisfying.There is large room to improve.Machine translation(MT)is the process of using modern computer system to translate text between natural language pairs,namely to translate text to another language,which is an important tool for cross-lingual communication in the information era.After decades of effort machine translation has improved much,the results have been improving,but the translation quality of machine translation systems is still not satisfactory for many applications.One crucial shortcoming lays in that the traditional phrase-based statistical machine translation could not perform much better.Alas,it takes long time to train the different models which are also quite complex.In this thesis we look into the discourse parsing and machine translation respectively.In discourse parsing respect,based on the idea of related works and ourselves,we make a straightforward yet efficient strategy to analyze the free text,to improve the results of every subtasks of discourse parser,such as the accuracy.In machine translation respect,we look into the state-ofthe-art machine translation system,namely neural machine translation(NMT),which is more effective than the traditional phrase-based statistical machine translation(PBSMT).Then,we introduce discourse parsing as features to NMT,to improve the performance of NMT.Experiments show that the results of both our two studies are improved,the translation quality of neural machine translation is improved based on the discourse parsing.