Retrieval-Augmented SCDG Understanding

The Mandatory Declassification Review (MDR) requirement of Executive Order 13526 has increased the administrative burden on those Government agencies that must manually review a significant volume of classified information each calendar year to ensure that no material is automatically declassified prior to review. In light of this challenge, Stottler Henke proposes to develop DART (Declassification Automation & Review Tools), an intelligent declassification review system that will more efficiently, consistently, and accurately determine the final disposition of historical records.  DART employs data-driven NLP models to understand security classification guides and reason about original markings.  These models combine generalized natural language understanding with task-specific capabilities related to security classification and declassification determinations.  DART will also rapidly find source documents and specific text portions containing the correct dispositions among very large document collections and archives.  Finally, user-friendly interfaces will streamline the DART declassification review process—drawing attention to specific portions that would benefit from manual review while at the same time helping those reviewers understand why a particular classification marking was suggested. This effort leverages our team’s past work and substantial experience in the areas of machine learning, sensitive information analysis, human-machine interaction design, and especially in natural language processing.