Fusion of Entity Information from Textual Data Sources (e.g. HUMINT)

The goal of this project is to push the boundaries of both research and development in entity fusion technologies through the creation of a robust, configurable, high-accuracy system for entity profile creation. The system will incorporate solutions that (1) lead to improved overall accuracy, increasing the recall of entity profile content without sacrificing precision through inclusion of information from non-name mentions and (2) provide realistic solutions to the issues of domain portability, user-sensitivity to noisy output, and processing speed. The key capability to be improved here is document-level coreference resolution with a focus on nominal and pronoun mentions of entities. We will pursue this improvement on three fronts. One is continued enhancement of the pairwise coreference resolution system through algorithmic changes to the statistical learning model. Two, we will develop a fast, accurate globally-optimal solution for merging mentions into entity profiles, which makes use of uncertainty in entity type tagging and pairwise-linking between entity mentions. Finally, we will develop an efficient method to utilize corpus-level word cooccurrence and predicate-argument statistics during document-level coreference. This information will assist in resolving the most difficult cases for coreference, the neutral pronouns `it' and `they' and nominal mentions.