Statistical Text Analysis with Training Semantics (STATS)

Charles River Analytics has developed an effective new technique for extracting complex information from Army intelligence reports. Grammar-assisted fine-tuning (GRAFT) grafts application- and source-specific linguistic knowledge onto a pretrained large language model (LLM). This enables the LLM to correctly interpret idiosyncratic language (e.g., military jargon and expressions) and to extract the specific semantic information needed by Army schemas and ontologies. Our research has shown that, without the GRAFT technique, LLMs perform poorly when extracting the required semantics from Army intelligence reports. With the GRAFT technique, the benefits of LLMs can be fully realized, including their deep knowledge of language and their rapid decision-making. The initial Phase II effort produced a prototype Statistical Text Analysis with Training Semantics (STATS) system featuring the GRAFT technique. We propose a sequential Phase II effort to mature the technology to Technical Readiness Level 6 (TRL 6), to perform demonstrations and evaluations on Sponsor data, and to prepare the STATS technology for integration into the Army Intelligence Data Platform (AIDP).