Graph Neural Network For Document-level Relation Extraction

Extracting the relationship between entities in a document is one of the basic tasks in the field of natural language processing.Document-level relation facts are contained in multiple sentences,compared with sentence-level relation facts,it has a complex way of interacting with entities.This thesis focuses on document-level relation extraction,and uses path reasoning to represent the complex interaction of entities.At the same time,for the imbalance and sparseness of relation labels in a large number of practical application scenarios,this thesis adopts the method of auxiliary learning to judge whether there is a relation between entities.The main contributions include:(1)Document-level relation extraction model based on multi-granularity entity feature representation: this thesis proposes a multi-granularity entity representation model based on graph convolutional network.The model uses heuristic rules to construct a document graph and uses graph convolutional network to transfer and aggregate the semantic information of the context,then capture the global feature representation of entities.Then,the model aggregates the path information between the target entity pairs through the attention mechanism and the interactive information between the target entity pairs is further mined.Through multi-granularity information fusion,the model explores the interactive information between entities deeper and improves the expression ability of entities features.(2)Relation extraction model with enhanced path probability: For the sparsity problem of relation label distribution,this model introduces auxiliary tasks that enhance the path probability between entities.The auxiliary task maximizes the path probability between pairs of related entities and minimizes the path probability between pairs of unrelated entities,making the model keep a higher focus on the entity pairs that have relations,and achieve the goal of reducing noise caused by the unrelated entity pairs to the graph model.(3)Document topic fusion hierarchical graph network cross-document relation extraction model: Aiming at the problem of target entity pairs appearing in different documents,this thesis proposes a hierarchical graph neural network model that integrates document topic information.The model hierarchically models the entity context information in a single document and the interaction information between entities across documents by constructing a two-layer graph neural network.Another,by combining with the document topic information,the model has preliminary cross-document relation extraction capabilities.This thesis focuses on the document-level relation extraction problem,and proposes improved models from different angle.This thesis conducts experiments on the corresponding datasets(Doc RED)and makes analysis.The experimental results show that the proposed model improves the performance of relation extraction and verify the effectiveness of the proposed model.