Customizable Text Extraction for Warfighters

Current state of the art information extraction (IE) systems are able to automatically extract specific types of information object such as entities, relationships, and events, that are known to be relevant to a particular domain. In order for these systems to achieve state of the art accuracies, however, it is typically necessary for human experts to spend much effort in tuning the system for the target domain, whether it be in annotating training data or writing grammars and lexicons. We propose to investigate techniques to significantly reduce the amount of effort required by human experts in order to port an IE system to a new domain. The focus of our efforts is Semantex\texttrademark, a state of the art IE system. In particular, we will explore various approaches including bootstrapping and active learning in order to ease training of entity detection, relationship detection, and event detection without incurring a huge penalty in accuracy.