Document Level Relation Extraction Based On Deep Learning

With the advent of the wave of globalization,informatization has led to an exponential increase in the information that people receive,and among them,text data is the largest.Text data usually exists in a structured form,and how to extract structured information from unstructured text information and use a computer for storage and processing is a problem to be solved in the current natural language processing application scenario.Relation extraction can extract structured triple information from unstructured text data,so the direction of entity relation extraction is the key direction to solve this problem.Although the research in the field of relation extraction in the field of natural language processing has made great progress in recent years,its current mainstream research direction is mainly based on sentence-level relation extraction.However,in real life,many relationship facts are hidden between sentences,and it is necessary to reason about multiple sentences to accurately extract the relationship facts between sentences.Therefore,the document-level relationship extraction direction is more worthy of study.At present,on the task of document-level relation extraction,the mainstream method is to generate a network graph,and then infer the nodes of this graph.But at present,the traditional graph neural network model used on the network graph can only learn the information of the surrounding nodes,and it is difficult to capture the information between the entity pairs,so its performance can be further improved.Based on the previous field of document-level relation extraction,an end-to-end document-level relation extraction model based on deep learning is proposed.The model includes coding module,image embedding module and classification module.The model can use the word preprocessing model to learn the context information of the word,and then construct a heterogeneous network graph according to the rules.By using the graph attention neural network model of the heterogeneous graph,the model is compared with the previous document-level relationship extraction model Able to better learn the facts about the relationships within and between sentences.And according to the characteristics of document-level relation extraction,a meta-path suitable for relation extraction within and between sentences is proposed based on heuristic rules.On the latest document-level relational data set,a comparative experiment with the current mainstream document-level relation extraction model proves the effectiveness of the model and the necessity of using heterogeneous network graphs,and then eliminates different meta-paths through ablation experiments.,Which proved the rationality of the designed meta-path.