SBIR Phase II: Customizable Question Answering System for Homeland Security and Commercial Applications

This Small Business Innovation Research (SBIR) Phase II project will result in a novel question-answering technology. The features of this technology are as follows: (1) Automatic filtering of questions. During Phase I, Language Computer Corporation (LCC) developed a system that decomposes high-level questions into low-level, fact-seeking questions. Some of these questions, however, turn out to be nonsensical. In Phase II, the firm proposes to submit all of the decomposed questions to a knowledge-based system, which will eliminate questions that are inconsistent with tacit knowledge. All of the questions that survive filtering will be passed back for processing by the question-answering system. (2) Aligning domain ontologies with a large reference ontology. During Phase I, LCC developed a tool that generates domain ontologies from raw text. During Phase II, the firm will extend this tool so that the domain ontologies are automatically aligned with an overarching domain-independent ontology. This alignment will permit deeper expansion of query concepts, because it will allow domain-independent concepts to be augmented with domain-dependent content. (3) Formal evaluation of semantic relations. The foundation of the question-answering system is semantic relations extracted from queries and documents. These relations will be evaluated to assess the relative contribution of each one to question answering. The result of this evaluation will establish which aspects of semantics are most useful to question- answering. This project will have a direct impact in the following areas: (1) The system can be deployed in commercial and government settings where the accuracy, coverage, reliability, and usability of the retrieved information are crucial. Ideal applications for the technology include homeland defense, CRM, education, medicine, and the law. (2) The system bridges the gap between domain-independent and domain-specific content. Domain ontologies are constructed automatically, and these ontologies are automatically aligned with a large reference ontology, so that queries can be simultaneously expanded into the terms appropriate to many different domains.