Research On Relation Extraction Method Of Graph Convolutional Neural Network For Multi-granularity Text

In recent years,research on relation extraction has received widespread attention and has been applied in many fields such as information retrieval,intelligent question answering and knowledge graphs.Relation extraction is an important field in natural language processing,which is to extract one or more relations from unstructured text.Different methods are usually used for relation extraction in different granularities of text,which can be divided into sentence-level relation extraction and document-level relation extraction.For sentence-level relation extraction,existing syntax-based dependency tree methods suffer from semantic inadequacy issues.Due to the adoption of overly aggressive or conservative pruning strategies,semantic information is insufficient or noise is excessive in the existing dependency tree-based relation extraction.For document-level relation extraction,information loss is easy to occur,due to the entities in the document are dispersed far apart.The low accuracy of document representation and document graph reasoning is also a major problem in the field of relation extraction.To solve the above problems,this thesis conducted in-depth research on both sentence-level and document-level relation extraction methods,proposing relation extraction methods for different granularities of text to overcome the shortcomings of existing methods.The main research work and innovations of this thesis are as follows:(1)For sentence-level relation extraction,this thesis proposes a Neural Attentional Relation Extraction Method with Dual Dependency Trees(DDT-REM).First,a dual dependency representation is constructed,including syntactic dependency representation and semantic dependency representation,and a syntactic dependency tree and a semantic dependency tree are constructed to obtain syntactic and semantic information of the sentence from both syntactic and semantic perspectives.Second,the local-global attention mechanism is used in the syntactic dependency tree to capture more detailed contextual semantic information,effectively avoiding the phenomenon of excessive or inadequate pruning.Finally,an extended graph convolutional neural network is designed for relation extraction,which is suitable for processing dependency structures,effectively improving the accuracy of relation extraction.(2)For document-level relation extraction,this thesis proposes a Document-Level Relation Extraction Method Based on Heterogeneous Graph Reasoning(HGR-DREM).First,BERT pre-training models are used to obtain document embedding representations,which are enriched by incorporating co-reference information and entity types.Second,a heterogenous graph is constructed for document-level relations,containing nodes and edges of different types representing the relations in the document.Then,the attention mechanism is used to calculate the relevance of nodes in the metapath and the attention is introduced into the adjacency matrix of the extended heterogenous graph.Finally,a graph convolutional neural network is used to obtain high-dimensional feature representation vectors,and the relation type between entities is determined through relation classification.The effectiveness of the proposed methods was verified through comparative experiments and ablation experiments on multiple public datasets.For sentence-level relation extraction,experiments on sentence length were conducted to verify the good generalization ability of the proposed method for sentences of different lengths,while for document-level relation extraction,attention performance experiments were conducted to find the appropriate number of attention heads to achieve the best prediction performance.