Graph Neural Network Relation Classification Combined With Commonsense Knowledge

At present,there are massive and multi-source unstructured text data on the network.How to quickly extract the content that users are interested in and useful is an urgent problem to be solved.Researchers have proposed the concept of using automated tools to transform unstructured data into structured data,and relation extraction is one of the most important steps.According to the marking degree of the dataset,relation extraction can be divided into supervised,semi-supervised and unsupervised.There are many algorithms for supervised relation extraction,that is,relation classification,which mainly focus on extracting the semantic features of the sentences to be classified by using deep neural network,and all of them have achieved good performance.With the emergence of advanced corpus resources such as knowledge graph,some researchers try to obtain the information related to the sample sentences in relation classification from knowledge graph in order to improve the performance of relation classification and enrich the characteristics of relation classification model.However,the application of knowledge graphs in relation classification tasks is still in its infancy and needs to be improved.For example :（1）most of the current relation classification methods only use the lexical features of knowledge graphs,ignoring the structural features of graphs;（2）The independent and complete relation classification model has not been constructed by fully utilizing the semantics of knowledge graph.（3）The current relation classification methods only use the relevant knowledge in the knowledge graph as a supplementary feature for the existing relation classification model,and have not realized the multi-directional integration of the semantics of the knowledge graph and the task of relation classification.In view of the above problems,this paper studies the relation classification method combined with external knowledge graph,and makes the following contributions:（1）The sample reconstruction of relation classification sentences based on knowledge graph is proposed.By reconstructing the samples,we construct the sub graph of the knowledge graph corresponding to the sentences to be classified,and capture the lexical features and graph structure features of the background context of the sentences in the knowledge graph;（2）Construct GCN-BLSTM-Attention network to extract the background symbolic features of knowledge subgraph.On this basis,a relation classification model DGCN（Dual Graph Convolutional Network）is proposed,which combines the text semantic features of the sentence itself and the symbolic features of the sub graph of the knowledge graph;（3）Further,KGCPR model is proposed,which is separated from the text semantics of the sentence itself and based only on the semantics of the knowledge graph.And the knowledge graph completion is used to complete the traditional sentence relation classification task;（4）Furthermore,combined with the semantics of knowledge graph and sentence text,a multi-level fusion relation classification model MLFRC（Multi-Level Fusion Relation Classification）based on attention mechanism is proposed to further improve the performance of relation classification.The experimental results show that the sample reconstruction based on knowledge graph can effectively capture the semantic and structural features of the sentences to be classified.Compared with the relation classification method using only lexical features,although DGCN model only carries out feature level fusion,it can also achieve better results.Furthermore,experimental results of the KGCPR model show that the relation classification method using only the semantics of the knowledge graph is feasible,and the relation completion task reconstructed from all samples on the knowledge graph is equivalent to the traditional sentence relation classification task.Finally,the experimental results of MLFRC model show that different semantics can guide and check each other,so as to further improve the performance of relation classification.