Adaptive Neural Machine Translation Solution OCR Research and Domain-Adaptation Scaling Prototype

This Subsequent Phase II project will extend and advance the innovation work done under the USAF Phase II SBIR with the National Air and Space Intelligence Center (NASIC).  The Office of Naval Intelligence (ONI) has three primary technical objectives for this project: Milestone 09 of NASIC’s SBIR effort integrated a basic/initial Optical Character Recognition  (OCR) capability with Lilt’s AI language translation software. While this insightful integration of two emerging technologies (OCR and AI language translation) significantly advanced the system’s overall functionality, ONI requires additional research and development to enhance this new capability and enable broader adoption of the tool that NASIC developed.  In this Navy Subsequent Phase II project, Lilt and ONI will a) research methods for enhanced OCR error detection capability, b) build software that enhances OCR functionality and quality, c) research methods for improved OCR of structured data, d) develop and implement methods to address OCR line break issues, e) improve OCR for bulk data translation, f) improve text extraction and PDF handling capabilities, and g) research methods for enhanced in-document OCR capability.  Milestone 07 of the NASIC SBIR effort involved retraining of the system on government data sources. In executing this milestone, Lilt and NASIC were able to successfully fine-tune Lilt’s translation algorithms, generating significant increases in machine translation quality. While this enabled the handful of translators at NASIC to benefit from domain-adapted translation, translators are greatly outnumbered by intelligence analysts at NASIC and ONI. Frequently, these intelligence analysts have little or no language capability. This technology enhancement presents an opportunity to scale domain-adapted translation across the NASIC/ONI analyst workforce, but this concept has never been implemented at scale at a USGOV translation center before. Effectively making use of this capability and transitioning this envisioned concept from theory to effective practice will require experimentation, innovation, research, and product development.  In this project, Lilt and ONI will a) develop and test a prototype of a system that can scale domain-adapted translation to ONI's analyst and watchstander workforce, b) investigate methods for tailored domain specialization and AI model optimization, c) quantify and assess the impacts of these advances, d) conduct research to enhance baseline model quality. Milestone 02 of the NASIC project involved the installation of Lilt on an IL2 commercial cloud environment; however, ONI, NASIC, and other IC components require IL4 GovCloud infrastructure to perform their intelligence missions.  In this project, Lilt and ONI will a) experiment with and report on emerging, innovative ATO pathways to rapidly deploy software in an IL4 environment and b) develop capabilities for overseas and deployed Navy units to securely access the software.