Research On Micro Discourse Nuclearity And Relation Recognition In Chinese

In recent years,with the focus of Natural Language Processing shifting from the word/sentence level to the discourse level,the discourse parsing task focusing on under-standing the overall semantics of texts is getting increased attention.Nuclearity recognition and relation recognition are two critical tasks in discourse parsing,which help to understand the context of the article by identifying the semantic relations and the nuclearity between discourse units.Most of the discourse parsing research focus on English,and that in Chinese is still in the infancy.Therefore,this dissertation mainly focuses on the identification meth-ods of discourse nuclearity and relations in Chinese,which includes the following three as-pects:(1)Discourse nuclearity recognition based on text matching methods.Most of the current research does not consider the language characteristics of the nu-clearity,therefore this dissertation proposes a text matching network for identifying the dis-course nuclearity.First,it captures both global dependency information and local n-gram information by combining Bi-LSTM and CNN in the encoder.Then,it introduces three com-ponents of text matching,the Cosine,Bilinear and Single Layer Network,to incorporate various similarities and interactions between discourse units.Finally,it makes the semantic match between the discourse unit and the paragraph to provide additional high-level seman-tic cues.Experimental results show that the proposed model outperforms various baselines.(2)Implicit discourse relation recognition based on sentence-level representations.Most of the neural network methods only simulate the single-pass reading process,while discourse relation recognition relies on the deep understanding of the text.Therefore,this dissertation proposes a three-layer attention network to simulate both the human bidi-rectional reading strategy and repeated reading process.First,it combines the self-attention model and the interactive attention model to capture the semantic connection between argu-ments in the text encoding stage.Then,it uses an attention layer with external memory to simulate the repeated reading process and generate the final argument representations.Ex-perimental results show that the proposed model outperforms various baselines.(3)Implicit discourse relation recognition based on topic-level representations.Most methods of discourse relation recognition rely on the sentence-level argument representations and do not perform well in Chinese.Therefore,this dissertation attempts to provide additional high-level discourse clues by introducing topic information,and proposes a topic tensor network based on gated convolutional network and simplified topic model.First,it uses a gated convolutional network encoder to learn the sentence-level argument representations.Then,it trains a simplified topic model through unsupervised learning to infer the latent topic distribution of arguments as the topic-level representations.Finally,it uses two factored tensor networks to model the interactions between arguments at both the sentence level and the topic level.In particular,tensor factorization is adopt to reduce the computational complexity of the model.Experimental results show that the proposed model outperforms various baselines.This dissertation proposes effective solutions on recognizing the discourse nuclearity and the discourse relations,which will provide a reference for further research of micro dis-course parsing in Chinese.