Information Extraction for New Emerging Noisy User-generated Microtext

This proposal presents an opportunity to address many challenges of information extraction and related techniques when applied to noisy user-generated microtext in PAI (publicly available information) and military “chat” data, in support of United States Air Force and Air National Guard intelligence analysts leveraging such sources in their missions.  Given the breadth of chat data platforms and dialects coupled with the need to address low resource settings, we propose a three-pronged approach: (i) focusing on conducting research into linguistically-informed decoders for deep networks, unsupervised decipherment of noisy text, and non-destructive text normalization; (ii) applying higher-order low-resource deep learning techniques to information extraction and entity resolution (as a component of information extraction), and (iii) using a hybrid approach of the two.  Evaluation approaches will consider the typical metrics of precision, recall and F-measure, as well as runtime performance and operational metrics including the effort required to develop and maintain training sets.  Software will be developed using agile methods, evaluated, and if found to be feasible, integrated into active operational transition paths in Intuidex’s Watchman for Defense (W4D) commercial product.