Customizable Text Extraction for Warfighters

Although information extraction (IE) systems are attractive because they automatically extract items of interest from free text such as entities, relationships, and events, development of accurate IE systems typically requires a great amount of labor by human experts. Semi-supervised approaches reduce the amount of labor, but contemporary versions of these approaches also hurt accuracy. The goal of this project is to develop superior semi-supervised approaches and to integrate them into an IE system. We will pursue improvement in several areas. First, we will explore the use of bootstrapping and active learning approaches for event detection. This includes investigating single-module versions of these approaches as well as pipeline-aware versions when event detection is integrated into the IE pipeline. Second, we will investigate the use of novel information sources in order to augment semi-supervised learning. These sources include the outputs from unsupervised methods such as topic modeling that extract structure from raw text. Third, we will investigate the use of domain adaptation techniques, which leverage human efforts in one domain in order to develop an IE system for another domain.