Knowledge Graph Time Reduction in Automated Curation Efforts (KG-TRACE)

Under this SBIR effort called Knowledge Graph Time Reduction in Automated Curation Efforts (KG-TRACE), Language Computer Corporation (LCC) proposes to develop a proof-of-concept system to research and assess the feasibility of alternative approaches (automatic and collaborative) to user-driven curation of a dynamic KG with the ultimate goal of drastically reducing the time required to correct automatically generated KG data. Under Phase I of this effort, we will focus on (1) developing support tools to allow analysts to more efficiently and effectively curate a dynamic KG (i.e., making changes and additions); (2) automatically identifying and highlighting information gaps in the KG; and (3) developing automated techniques to extrapolate from user-driven modifications to trigger or suggest similar changes throughout the KG. As a part of this effort, we will apply a suite of evaluation metrics to estimate the speed, accuracy, and usability characteristics of the system in order to show the efficacy and relative merits of our approaches in preparation for Phase II. Over the course of this project, LCC will build upon our unique combination of expertise in a variety of diverse technical areas including Knowledge Base Population (KBP), Information Extraction (IE), Cross-Document Coreference (CDECR), Question Answering (QA), logical representation and inference, and analyst tool development. We will leverage this expertise (along with our experience with AF and DoD analytic workflows) to support the diverse set of applications supported by knowledge graphs, including QA, information retrieval, pattern-of-life-analysis, and threat detection. As a result of this Phase I SBIR, we will have demonstrated the feasibility of AI-assisted, user-driven curation of a dynamic KG and will have identified additional challenges to be addressed and enhancements to be incorporated into a full-scale software prototype under Phase II. This software will be flexible and will be designed to operate over a variety of KG types with the goal of reducing the curation time costs associated with the use of dynamic KG data while increasing overall analyst confidence in KG data quality.