Probabilistic Knowledge Graph Construction: Compositional and Incremental Approaches

Knowledge base construction consists of two tasks: extracting information from external sources (knowledge population), and inferring missing information through a statistical analysis on the extracted information (knowledge completion). In many cases, there are not enough external sources to extract information for a comprehensive knowledge base, hence the need for a separate completion step. An incremental knowledge population approach via labelling of human experts can help to reduce the gap between these two processes. We propose a new probabilistic knowledge base factorisation method that benefits from the path structure of existing knowledge (e.g. syllogism). Our method enables a common computation and modelling approach to be used for both knowledge base construction tasks. Empirical experiments show that our model improves over the non-probabilistic and non-path counterparts on the knowledge completion task. The probabilistic formulation allows us to develop an incremental knowledge population model that trades off exploitation and exploration. We show that our proposed approach for incremental knowledge population performs significantly better than sampling knowledge triples at random. However the path structure does not seem to help in this setting. In applications where there is a continual improvement in knowledge bases, our common probabilistic model bridges the gap between knowledge population and knowledge completion.