Algorithm Research And Implementation Of Document-Level Relation Extraction

Document-level relation extraction aims to extract the relations between all entity pairs in a document.Compared to traditional sentencelevel relation extraction,it faces more challenges.In recent years,document-level relation extraction based on deep learning neural network has become the mainstream research direction.However,there are still some problems to be solved in document-level relation extraction.First of all,researchers do not pay enough attention to the key information of full text.Second,the problem of multiple mentions of the same entity that is unique to document-level relation extraction is under-resolved.In addition,for the acquisition of reasoning ability,most researches focus on the method based on document graph,which needs to build document-level graph structure by manual method,requires expert ability and the training speed is often slow.Researchers do not pay enough attention to the key information of the overall information of the full text.At the same time,the problem of multiple mentions of entity,which is unique to document-level relation extraction,is not sufficiently solved.In addition,for the acquisition of reasoning ability,most researches focus on the method based on document graph,which needs to build document-level graph structure by manual method,requires expert ability and the training speed is often slow.In view of the above problems,the main research work and contributions of this paper include:(1)This paper proposes an entity-pair representation enhancement module based on explicitly fusing full-text information and mention contribution weights.This paper focuses on considering that the text carrier of document relation extraction is the document,so a method based on the attention mechanism is proposed to closely integrate the document information with the head and tail entity information,and potentially learn the unique information of the document,so as to achieve the purpose of enhancing the relation classification effect.At the same time,this paper proposes a method to calculate the contribution weight of mentions,which can dynamically calculate the contribution weight of each mention for entity pairs.(2)This paper introduces a cross-attention module and evidence sentence joint training paradigm.The former can only use the self-attention mechanism without the method of document graph to enable entity pairs to have reasoning ability.The latter takes the evidence sentence extraction task as a subtask to construct a joint training paradigm.(3)This paper conducts sufficient experiments on the public dataset.It empirically shows that the method proposed in this paper has a certain effect compared with the previous model,which proves the effectiveness of the method proposed in this paper.