Research On Key Technologies Of Chinese Coreference Resolution Based On Deep Learning

Linguistically,in order to avoid the problems of redundant sentence structure and semantic ambiguity caused by the repetition of the same word,the general method is to replace the existing pronouns or nouns with anaphors to simplify sentence expression.However,it will lead to the confusion of reference.Therefore,coreference resolution task play an important role in the field of semantic understanding in NLP.Its main task is to build a one-to-one mapping relationship between anaphors and antecedents automatically by integrating prior language knowledge,so as to avoid semantic understanding deviation and low weight of word modeling.At present,plenty of work has been done on coreference resolution.With the rise of deep learning technology,the research of coreference resolution based on deep learning framework has gradually become the mainstream research method,which focuses on how to use phrase-based word embedding information in sentences to effectively identify the entities and the relationship between them.But,there exists a main problems: In order to improve the performance of coreference resolution task,the effective fusion of multi-level semantic information,such as char level,word level and sentence level,has not been fully considered,especially Chinese coreference resolution task.Chinese is difference from Latin,which has unique language characteristics,therefor the Chinese coreference resolution task is more challenging.In thesis,we focus on how to build an effective hierarchical coreference resolution model for the integration of external knowledge,including the following aspects: At character level,rich part of speech tagging information is mainly considered;At word level,word encoding feature and phrase-level feature information based on convolutional neural network are in consideration;At sentence level,using the local context information of the sentence where the candidate entity is located,so as to effectively improve the accuracy of inter entity relationship judgment.Experiments on the Chinese coreference resolution dataset Onto Notes5.0 show that applying Bert-wwm-ext and multi-layer feature fusion to the Chinese coreference resolution model achieves strong improvement(+3.21 F1)compared with the existing SOTA(state of the art)method.