Theia Automated Data Labeling and Curation

The Army intelligence community seeks an artificial intelligence (AI)-based automated and reliable data labeling and curation system that allows users to quickly search for, filter, and select datasets for further downstream analysis. Through a Phase I Small Business Innovative Research (SBIR) effort, Illumination Works (ILW) proved the feasibility of Theia, our AI-driven solution for automated data labeling and curation. ILWÆs Theia has three core components. The AI Processing Engine applies AI/machine learning (ML) techniques from computer vision and natural language processing (NLP) to automatically label intelligence data by identifying and segmenting key entities. Labels are curated into enhanced metadata packages that include higher-order categorical summaries, entity labels, and entity segment extraction metrics. The Knowledge Store will store entity labels and higher-order topics/classes identified and will further enrich data with relevance scoring, thus providing improved curation and automated label validation. The User Interface (UI) provides human-machine teaming, giving users an interface to query and locate relevant data, enabling analysts to filter out the data noise and focus attention on informative data for the question at hand. The Phase II effort enhances TheiaÆs capabilities to a TRL 6 state where ILW will complete the build out of a production-ready prototype to be inserted into an Army environment of choice in Phase III. The key differentiator in TheiaÆs implementation of autonomous labeling is in the ability to programmatically traverse intelligence documents and data to precisely organize all entities and relationships within analyst feeds (reports, articles, documents), to construct a knowledge retrieval system beyond a simple keyword search or phrase search engine. Using the technology developed within Theia during Phase II, users will have the ability to utilize labels for their own modeling and analysis. Labeling will be grown to include self-learning, semi-supervised methods and will be conducted over a plethora of intelligence sources including reporting, imagery, tabular data, and more. To assist analysts in using these labels, TheiaÆs UI will be developed into a usable prototype with a focus on retrieving documents and labels from the underlying label database.ĀĀ???????