Knowledge Graph Completion Based on Graph Representation and Probability Model

Knowledge Graph provides an effective scheme for unstructured knowledge on the Internet. but, to a large extent, the lack of information in the knowledge graph restricts the application of the knowledge graph. Therefore, knowledge graph completion has been an attractive research area. In recent years, many researchers apply graph convolutional neural networks to knowledge graph embedding so as to solve this problem, including R-GCN, SACN et al. The current state-of-the-art SACN uses graph convolutional neural as an encoder to make more accurate embeddings of graph nodes, and uses a convolutional network as a decoder, leading a good performance of link prediction task. In this work, we produce a novel graph representation model based on SACN---property graph convolutional network called PGCN. PGCN treats the knowledge graph as a property graph, regarding the initial embedding vector of entities and relations as the property of nodes and edges. What makes the model different is that it introduces the node clustering before convolving the nodes, so that the graph takes into account the importance of neighbors when aggregating the neighbors of nodes. We adopt the probability model Conv-TransE proposed by SACN for the modeling of relation of graph, Conv-TransE takes advantages of ConvE and use probability method which greatly improves the efficiency of the experiment and avoids the construction of the corrupted triplet. We conduct experiments on the standard datasets WN18RR and FB15k-237, and demonstrate that our new model achieves better performance.