Coreference Resolution Model Based On Bidirectional And Multidimensional Attention Mechanism

With the development of science and the practical needs of industrial application,it is urgent to understand the text at a deeper level.In recent years,with the deepening of the research task of natural language processing,the technology of natural language processing for short text is becoming more and more mature.As one of the key tasks of natural language processing,coreference resolution has been widely concerned by scholars at home and abroad.This paper aims at the task of coreference resolution at the text level,that is,for any two mentions in the same document,if they point to the same entity in the real world,the two mentions are said to have coreference relationship.When multiple mentions point to the same entity,the multiple mentions form a coreference chain,and any two mentions have coreference relationship in the same coreference chain.At present,among many existing common the coreference resolution models,Lee's the coreference resolution model greatly reduces manual workload,and at the same time uses end-to-end neural network technology to effectively improve the performance of the coreference resolution model.This method makes this model has become one of the classic models of the coreference resolution model without using additional tools.In this paper,by analyzing the advantages and characteristics of this model in coreference resolution task,and comparing with other coreference resolution models,this model is finally selected as the baseline model of this paper.The task of this paper is to construct a coreference resolution model.in English texts,and to mine the coreference relationship in documents by capturing the meaning and context of words,so as to help machines understand and analyze texts more effectively.The thesis combines deep learning theory and natural language processing technology,starting from the linguistic characteristics of the data set,and taking the end-to-end coreference resolution model as the baseline model to explore the relationship between expressions from word types,word features and context features,and analyzes the scheme that can be improved.An improved model based on end-toend coreference is proposed,that is,unitary features are added to reduce data sparsity.By referring to the Bert pre-trained language model with stronger feature extraction capability,it can provide the encoder with better vector representation and obtain vectors with richer semantic knowledge,so as to solve the polysemy.At the same time,the score function of the candidate antecedent and the alignment function of the attention mechanism are redesigned to obtain more semantic information and improve the performance of the text level coreference resolution model.In this paper,a coreference resolution model based on bidirectional multidimensional attention mechanism is proposed.bidirectional LSTM and self-attention mechanism are used to fully capture the semantic dependency between phrases and generate sentence representation.The mechanism of bidirectional multi-dimensional attention is used to capture the semantic features between sentences and generate document representation,so as to give full consideration to the local and global features in the text and construct a fine and accurate coreference chain.The effects of LSTM layer number,pretraining language model selection and unary feature addition on the performance of coreference resolution model are discussed through a number of experiments.It is proved through experiments that the model proposed in this paper has improved efficiency and performance compared with the baseline model and the current more classical coreference resolution model,among which the accuracy and recall rate are significantly improved.