Research On Relation Inference In The Open Knowledge Network

In the era of big data network, not only more and more data is needed, but also useful knowledge from the data is increasingly demanded. Relation inference based on large-scale knowledge network uses the knowledge networks to dig out hidden relationships between entities in favor of network applications such as information r etrieval, information recommendation and intelligent question and answer. Depending on whether entity relationships change over time, relation inference can be divided into static relation inference and dynamic relation inference. Dynamic relation inference can be further divided into non-sequential dynamic relation inference and sequential dynamic relation inference.The former only infers the existence of relationships, while the latter can identify their generation time at the same time. The technical difficulties in traditional entity relation inference can be attributed to correlative representation and comprehensive utilization of the complex heterogeneous information, such as attribute, structure and time. However, existing knowledge network models are difficult to solve the problem effectively. Therefore, new knowledge network model and relation inference are studied in this dissertation. Research contributions are listed as follows:(1) Knowledge network model is researched. For traditional knowledge network model can not fully describe complex heterogeneous information, Open Knowledge Network(Open KN) is proposed based on an eight-tuple model. Open Knowledge Network is a heterogeneous network, whose nodes and edges come with time, space, property, and a series of functions or operators. Compared with the traditional knowledge network, Open Knowledge Network is an open, adaptive, evolutionable heterogeneous network, which can do favor to deep excavation of information.(2) Static relation inference is studied. For traditional methods cannot fully characterize relationships with simple structure information used, a static relation inference method based on Open Knowledge Network is presented. In this method the property is selected from the rich property information in the Open KN through a decision tree. Then, the feature value of relation path is computed by using the idea of random walk. Finally, relation inference is achieved by integrating property and structure information. Experiment results show that the new method improves the inference accuracy compared with traditional methods.(3)Non-sequential dynamic relation inference is explored. While the inference complexity of existing supervised inference methods is high, unsupervised inference methods have low inference accuracy. A non-sequential relationship inference method based on Open KN is put forward, which takes full advantage of the time information in Open KN. Time information is integrated into the hybrid knapsack problem, and meaningful link extension patterns are selected by solving the knapsack problem. Finally, the inference results are obtained by matching patterns. Experiment results show that the new method achieves a higher accuracy ratio, and it is an unsupervised inference method with low complexity for big data environment.(4) Sequential dynamic relation inference is researched. For existing methods can not adequately express the inherent connection between time and structure, a sequential dynamic relation inference method based on OpenKN is designed. Firstly, the time information is integrated into the structure of the network. Then the relationship and its generation time is inferred by using the logistic regression model. Experiments show that the new method improves inference performance.(5) Though combining key technical achievements with users’ demands, a relation inference system based on Open KN is developed. The system mainly consists of three modules: Open Knowledge Network underlying storage module, offline model training and patterns selection module, online relation inference module. Entity relation inference is realized and put into use in a number of workplaces.