Knowledge Representation And Inference On Internet Data

Knowledge representation and inference is the field of artificial intelligence dedicated to rep-resenting information about the world in a form that a computer system can utilize to solve com-plex tasks. Knowledge representation incorporates findings about how humans solve problems and represent knowledge in order to design formalisms that will make complex systems easier to design and build. Knowledge inference incorporates findings from logic to automate various kinds of reasoning. The two processes are mutually reinforced.The various types of Internet data are mainly divided into two categories:the static and struc-tured data, and the dynamic and unstructured data. For the static and structured data, knowledge representation mainly aims to convert the sign-based data into computable vectors, and knowledge inference helps to complement the incomplete data. And for the dynamic and unstructured data, knowledge representation aims to build the relational structure of the data set, while knowledge inference purposes to learn the relational patterns of the data and discover important and inter-esting information. Based on the above observations, this thesis aims to provide algorithms of knowledge representation and inference for the both categories of data, in order to take advantage of the Internet knowledge. Consequently, this thesis studies the knowledge representation and inference problem and has the following contributions:First, we focus on the knowledge graph embedding problem with the static and structured data. This thesis proposes a bilinear learning framework which performs cross-entity knowledge relation analysis in the continuous vector space (derived from knowledge embedding). In the framework, we effectively model the intrinsic correlations among different types of knowledge relations within a max-margin multi-relational ranking scheme, which jointly optimizes the tasks of entity embedding and cross-entity relation prediction in terms of multi-relational structures of the knowledge graph.In order to represent and infer the high-order semantic information on the knowledge graph, this thesis proposes a joint embedding scheme, which simultaneously fulfills the two different but correlated embedding tasks (i.e., entity embedding and relation embedding). In the embedding scheme, the explicit representations of both entities and relations are learned by involving the high-order contextual information of the knowledge base, which is capable of effectively capturing the intrinsic topological structures in the learned embedding spaces.On the dynamic and unstructured data, we solve the representative problem of tracking news article evolution. This thesis proposes a context-dependent news knowledge discovery method based on temporally-successive news article connection using subgraph learning. The proposed method is able to adaptively construct a cross-article link network along the temporal dimension, and effectively discovers the news event pattern by dense subgraph learning to use the contextual news connection structures. Based on the learning structures, fast and accurate link path inference methods are presented.