Multi-step Reasoning for Knowledge Graph Completion

摘要

Most Knowledge Graphs (KGs) have the problem of missing relationships, although embedding methods that map entities and relationships to low-dimensional spaces can achieve certain results. But they only focus on the direct relationship between entities and ignore the importance of the semantic features of entity types. In addition, a large amount of unstructured text can effectively supplement the semantic information of the Knowledge Graph (KG). In order to solve the problem of scattered knowledge and lack of relationships in the KG, we propose a neural network model which is based on attention mechanism. The model combines the semantic information and the entity type captured by the attention layer to predict the link of tripe. We performed link prediction tasks on public datasets, and under the premise of sufficient data sources, also confirmed that our method is innovative and has research value.