Research On Relation Extraction Based On Graph Convolutions

With the development of information technology and the popularization of the Internet,there are mass of information been generated and widely spread on the Internet.Because of the reduction of the cost of information generation and dissemination,the convenience of people to obtain information has greatly improved.However,the subsequent information explosion problem has started to affect the efficiency of people's information retrieval.How to obtain the information users need form the amount of information on the Internet quickly and accurately has become an urgent problem.In order to solve this problem,information extraction technology has emerged as the times require,and has attracted widespread attention from research scholars in recent years.Relation extraction is one of the most important tasks in the field of information extraction.Existing relation extraction approaches mainly rely on exploiting external resources and background knowledge to improve the performance and ignore the correlations between entity pairs which are helpful for relation extraction.We present the concept of entity pair graph to represent the correlations between entity pairs and proposed a novel entity pairs graph based neural network model,relying on graph convolutional network to capture the topological features of an entity pair graph.EPGNN combines sentence semantic features generated by pre-trained BERT model with graph topological features for relation extraction.We proposed an entity association graph and attention based model for relation extraction.Considering that the information of the related entities in the corpus can assist the relation extraction of the target entity,we proposed the concept of entity association graph.An entity association graph constructed according to the co-occurrence relation between entities.The weights of the related entities are calculated by the attention mechanism.Then the features of associated entities are formed by the weighted sum of the features of all related entities.The combination of self-attention and Bi-GRU is used to extract sentence semantic features.Combining the features of associated entities and sentence semantic features for relation extraction.Based on the ERANN model,we proposed an entity pair graph based model(EPGNN).For the ERANN model,the independent entity is the object,resulting in the loss of the entity's overall association,we take entity pairs as nodes and common entities as edges to construct entity pair graph.In view of the transitivity of the relationship between entities,it is not comprehensive enough to consider only the directly related entities of the entity.The ERANN model is limited by its own characteristics of the attention mechanism,and is powerless for multi-level entity association relationships.To overcome this deficiency,EPGNN uses a multi-layer graph convolutional network to capture the topological features of the target entity pair.A pre-trained language model BERT is used to extract the sentence semantic features.EPGNN combines graph topological features and sentence semantic features for relation extraction.To verify the performance of our proposed model,the current mainstream relation extraction methods are used as benchmarks,and comparative experiments and results analysis are performed with the proposed ERANN model and EPGNN model on two public data sets commonly used in the fields of relation extraction.The ERANN model fully exploits the association relationship of target entity pairs by constructing entity association graphs,and then complements the global characteristics of target entity pairs based on sentence semantic features,and achieves better results than existing methods.The improved EPGNN model of feature construction method can more effectively capture the topological features and text semantic features of multi-order entities,and has achieved the current optimal relationship extraction effect on two public data sets.