Research On Knowledge Graph Automatic Evolution

Knowledge Graphs (KGs) are important structured semantic knowl-edge databases, consisting of a vast amount of knowledge facts, each of which is formed as a triple like (subject entity, predicate relationship, ob-ject entity). At present, KGs have become core data which supports var-ious artificial intelligence applications such as semantic searching, ques-tion answering, etc. Thus, it is very valuable to research on KG. The data source of KG includes encyclopedia data, un-structured text data and searching log, etc. Automatic knowledge graph construction (AKGC)aims to expand KG via continuously adding new knowledge facts to KG.However, AKGC is a never-ending dynamic process since the knowledge facts in the real world are always increasing. Thus, KGs always suffer from incompleteness. In the dissertation, we aim to reason/predict the implied/missing knowledge facts by exploring the existing knowledge facts in KG, and then expand the knowledge coverage of KG, which is called the evolution of KG. This research focuses on the issue on knowledge graph automatic evolution.First, to solve the incompleteness problem of KGs, we propose a pairwise-interaction differentiated embedding model (PIDE), which is based on the following two assumptions: 1.The confidence of the knowledge fact(subject entity, predicate relationship, object entity) would be determined by their pairwise interactions (i.e. (subject entity, object entity), (sub-ject entity, predicate relationship) and (object entity, predicate relation-ship)); 2.Entities have semantic and syntactic information, and predicate relationships have syntactic information. Our approach learns latent se-mantic and syntax representations of entities and relationships by training PIDE model, and then predict implied/missing knowledge facts in KG, in order to manage knowledge graph automatic evolution. In addition, we also propose a maximum ranking likelihood-based optimization algorithm.Empirical experiments verify the good performance of the proposed model and the optimization algorithm.Second, the PIDE-based approach is unable to predict missing knowl-edge fact which contains out-of-KG entities due to lack of latent represen-tations of out-of-KG entities. To solve the problem, we consider to use auxiliary text corpus. In general, KGs include entities along with their text descriptions, which explain the corresponding entities' meaning in detail and possess rich information. Motivated by the zero-shot learning,we propose a novel model, namely JointE, jointly learning embeddings from KG and entity descriptions, to extend KG by adding new knowledge facts with out-of-KG entities. The key idea of the model is that the latent representations of out-of-KG entities can be calculated based on their de-scriptions. Empirical comparisons on benchmark datasets show that the proposed JointE model outperforms state-of-the-art approaches.However, a few KGs do not provide entities' text descriptions. In this situation, the JointE model should be unavailable. At present, the increasing unstructured text corpus provides people with a large volume of information about entities. Based on the information of entities, we propose a Translating Embedding Neural Network model, which is trained on knowledge graph and unstructured text corpus, trying to learn new entities from unstructured text corpus and to expand KG. Experimental results demonstrate the excellent performance of our proposed algorithm.Motivated by the algorithms of knowledge facts prediction in KGs,we propose a embedding-based algorithm to model information recom-mendation data. Information recommendation data, consists of lots of triples (<user, rating, item>), is similar with knowledge graph in terms of the inner structure. The algorithm treats the data as multi-relational network just like knowledge graph, manages rating prediction, user predic-tion and item prediction simultaneously in recommendation systems. We also apply the proposed model for cross-domain recommendation, which is able to realize recommendation generation in multiple domains. Empirical comparison validates the effectiveness of the proposed model.