Knowledge Graph (KG) usually contains billions of facts about the real world, where a fact is represented as a triplet in the form of (head entity, relation, tail entity). KG is a complex network and consists of numerous nodes (entities) and edges (relations). Given that most KGs are noisy and far from being complete, KG analysis and completion methods are becoming more and more important. Knowledge graph embedding (KGE) aims to embed entities and relations in a low dimensional and continuous vector space, which is proven to be a quite efficient and effective method in knowledge graph completion tasks. KGE models devise various kinds of score functions to evaluate each fact in KG, which assign high points for true facts and low points for invalid ones. In a KG of the real world, some nodes may have hundreds of links with other nodes. There is a wealth of information around an entity, and the surrounding information (i.e., the sub-graph structure information) of one entity can make a significant contribution to predicting new facts. However, many previous works including, translational approaches such as Trans(E, H, R, and D), factorization approaches such as DistMult, ComplEx, and other deep learning approaches such as NTN, ConvE, concentrate on rating each fact in an isolated and separated way and lack a specially designed mechanism to learn the sub-graph structure information of the entity in KG. To conquer this challenge, we leverage the information fusion mechanism (Graph2Seq) used in graph neural network which is specially designed for graph-structured data, to learn fusion embeddings for entities in KG. And a novel fusion embedding learning KGE model (referred as G2SKGE) which aims to learn the sub-graph structure information of the entity in KG is proposed. With empirical experiments on four benchmark datasets, our proposed model achieves promising results and outperforms the state-of-the-art models.

With the rapid development and widespread application of Knowledge graphs (KGs) in many artificial intelligence tasks, a large number of efforts have been made to refine them and increase their quality. Knowedge graph embedding (KGE) has become one of the main refinement tasks, which aims to predict missing facts based on existing ones in KGs. However, there are still mainly two difficult unresolved challenges: (i) how to leverage the local structural features of entities and the potential soft logical rules to learn more expressive embedding of entites and relations; and (ii) how to combine these two learning processes into one unified model. To conquer these problems, we propose a novel KGE model named JSSKGE, which can \textbf{J}ointly learn the local \textbf{S}tructural features of entities and \textbf{S}oft logical rules. Firstly, we employ graph attention networks which are specially designed for graph-structured data to aggregate the local structural information of nodes. Then, we utilize soft logical rules implicated in KGs as an expert to further rectify the embeddings of entities and relations. By jointly learning, we can obtain more informative embeddings to predict new facts. With experiments on four commonly used datasets, the JSSKGE obtains better performance than state-of-the-art approaches.

The knowledge graph plays an important role in the construction of artificial intelligence applications. However, the incompleteness of the knowledge graph seriously affects the performance of downstream applications. The problem has fueled a lot of researches on knowledge graph completion (also known as the tasks of link prediction). Reinforcement learning-based multi-hop reasoning that formulates link prediction as a sequential decision problem has also become an interesting and promising approach. Nevertheless, in an incomplete knowledge graph environment, the policy-based agent might travel a large number of low-quality or spurious search trajectories, which inhibits the model performance. Therefore, in this paper, we propose a path and structural features-enhanced reinforcement learning model (referred as PGATRL). First, we leverage the path constraint resource allocation algorithm to mine high-quality inference paths, which can be employed to pre-train the LSTM path encoder module in the reinforcement learning architecture, and thus play a role in guiding the agent’s action decision-making. Second, we exploit the adapted graph attention networks to encode the local structural features of an entity, which can provide more evidence for the agent to find a more suitable reasoning path. With extensive experiments on several benchmark datasets, our proposed approach gains significant improvements compared with the state-of-the-art baselines.