Deduction for Natural Language Access to Data

Introduction

We outline a general approach to automated natural-language question answering that uses first-order logic and automated deduction. Our interest is in answering queries over structured data resources. We are concerned with queries whose answer is not stored directly in a single database but rather must be deduced and computed from information provided by a number of resources, which may not have been designed to work together. While the obstacles to understanding natural language queries are formidable, we simplify the problem by limiting ourselves to a well-understood subject domain and a known set of data resources. Using domain knowledge, queries in natural language are mapped to a logical representation and interpreted using an automated reasoner over a logical theory with semantic links to target knowledge sources.
Examples are drawn from a prototype system called Quest, which is being developed for a business enterprise question-answering application. Users of such a system can query complex databases without needing to know the structure of the target knowledge sources or to write programs in a database query language.