Knowledge Graph Completion With Pattern-Based Methods

Knowledge graphs (KGs) are popularly used to develop several intelligent applications. Revealing valuable knowledge hidden in these graphs opened up a branch of research, known as KG reasoning, aiming at predicting the missing links. Some methods take advantage of external information such as entity description but at the cost of more computational complexity. Besides, most of the current techniques focus solely on local information in the KG. However, valuable global information in the entire graph can be utilized in the learning process. In this paper, we propose a Pattern-based Knowledge Graph Completion (PKGC) method that consists of three phases. The first phase utilizes multi-source information and expands the KG using entity description as external information with efficient Natural Language Processing (NLP) techniques. In the second phase, we mine frequent patterns from the expanded KG, extract connections between them and assign entities to the patterns that construct the abstraction layer. Based on the extracted patterns, connections, and entity assignments, a flow network is constructed on the abstraction layer in the third phase. We use global internal information, namely patterns, by adapting the minimum-cost circulation problem to the flow network. This way the links in a larger neighborhood are involved in the inference. We conducted experiments on the link prediction task and evaluated the training time on two benchmark datasets, WordNet and Freebase. Experiments have demonstrated that the proposed method is superior to the state-of-the-art methods and that pattern extraction is effective for knowledge graph completion tasks.