High-Precision Agile Active Learning for Domain-Customizable Information Extraction (HALCYON)

The dynamic operational environments in which Air Force users operate today require information extraction systems that can be rapidly – and easily – customized to new and challenging domains. In order to address operational demands for textual information, Language Computer Corporation (LCC) has developed a customizable information extraction system, known as CiceroCustom, which enables military and intelligence personnel to extract information from sources of unstructured textual information (including OSINT and HUMINT) quickly and efficiently. In this Phase I SBIR effort, called High-Precision Agile Active Learning for Domain-Customizable Information Extraction (HALCYON), LCC will extend the customizable information extraction capacity provided by CiceroCustom with a new framework which can be used to enhance the quality and accuracy of domain customizations performed by users. We plan to build an enhanced prototype which incorporates (1) an agile customization framework which leverages a novel paradigm for active learning, (2) a context-driven mechanism for customizing extractors to specific domains that allows for the incorporation of diverse forms of user input, (3) a novel method for integrating domain-specific knowledge into an information extraction system, and (4) a robust textual reasoning capability which leverages a state-of-the-art textual entailment system in order to reason about domain knowledge for extraction.