Using Rich Inference to Find Novel Answers to Questions

The Web is continuously enriched with data and has become a large knowledge repository. However, machines are unable to fully exploit this vast knowledge space in performing reasoning tasks such as question answering. This inability limits the extent of inference and ultimately limits the range of questions they can answer. We argue that the quality and range of answers generated by a question-answering system is significantly improved when we use rich reasoning techniques to infer novel knowledge from web data. By finding and aggregating facts from different knowledge bases, an agent can obtain a better representation of a domain and hence infer new facts which did not exist in any of the original knowledge sources. We intend to explore rich semantic representations and rich forms of reasoning. These include the curation of data and the use of a combination of heuristics, logic and probabilistic techniques to infer answers. This approach will minimize noise and uncertainty in the knowledge for reasoning. Our customized representations will suit the problem to be solved rather than being restricted by the formalisms used in the sources. We plan to implement this in a question-answering system that exploits a vast
set of knowledge bases such as ontologies and Linked Data repositories. Our question-answering system will focus on questions which require rich inferences such as prediction and composition of answers from several pieces of information.