Practical Probabilistic Reasoning in Web Knowledge Graphs

点击翻译按钮获取中文摘要

The generation of so-called 'Knowledge Graphs' from heterogeneous Web sources has recently attracted significant attention both by academia and industry and a number of large, highly visible knowledge graphs have been created by research groups and large search engine providers such as Google and Microsoft. Despite significant improvements in the performance of the extraction methods used, these knowledge graphs still contain a number of inconsistencies that are caused by incorrect information generation through the extraction. The proposed project will investigate methods for checking the consistency of real world knowledge graphs that have been extracted from heteriogeneous sources. Classical, logic-based methods for consistency checking have serious limitations in this scenrio as they fail to address a number of key issues including: (1) probabilistic measures of certainty for extracted facts that are often provided by teh extraction method, (2) numerical attributes and constraints of objects described in the knowledge graph and (3) temporal information about the validity of facts. In our previous work, we have already developed methods that take probabilistic certainty of facts into account by designing a template language for Markov-Logic Networks that simulates the logical semantics of knowledge graphs and allows us to reduce consistency checking in knowledge graphs to MAP-State inference in Markov Logic. In the course of the proposed project, we want to extend this work to support numerical attributes and constraints as well as temporal reasoning to better address the challenges of consistency chekcing in real world knowledge graphs.As a first step, we will extend Markov-Logic Networks with numerical attributes and -constraints and develop efficient methods for reasoning in the extended Markov-Logic model. Based on this fundamental extension, we will develop an expressive template language for this extended model that allows us to encode numerical and temporal information in addition to the logical model and provides a reduction of consistency chekcing to MAP-State inference in the extended Markov-Logic model. Finally, we want to systematically evaluate the template language and the underlying methods on the basis of well known knowledge graphs, i.e. DBpedia, YAGO and NELL. In a second shorter iteration of the project, the methods and the language will be improved based on the results of this evaluation. Besides scientific publications, concrete results of the project will be an extension of the RockIT Reasoner for Markov Logic Networks and an extension of its existing online interface to map existing knowledge graphs to the template language developed in the proejct.