A context mechanism for an inference-based question answering system

Question answering (QA) techniques have predominantly focused on improving semantic parsing and information retrieval steps. Recent work has seen significant advances using deep neural networks to tackle these problems. However, not much emphasis has been put on incorporating contextual information into the QA process. More so in inference-based QA methods where, in addition to information retrieval (IR), there is the need for a non-deterministic composition of different operations on data from diverse sources.

In this paper, we formalise the idea of context and describe how it can be injected into a question answering process which, in addition to the retrieval of facts, requires the use of deduction, statistical and mathematical operations. We refer to this as an inference-based QA process. We show how this can improve the answers returned by constraining the key operations in the QA pipeline to contextual information. Context includes a user's environment and preferences such as how they might want to trade off accuracy over speed in the inference process. The latter informs the choice of inference methods that are used to answer the question. We explore these ideas using an inference-based QA framework that draws on structured data from diverse knowledge graphs, including commonsense knowledge found in sources such as Wikidata, decomposes questions recursively and combines retrieved facts using arithmetic and statistical operations, including making predictions. Experiments on questions based on Wikidata and the World Bank Open Data set validates the effectiveness of the proposed approach.

Our primary contribution is our approach to incorporating context information in the QA process, especially when inferring answers that cannot be found by traditional IR methods.