Advanced Time-Stamping of Events from Unstructured Text for Battlespace Awareness

Current state-of-the-art in text extraction technology can associate a time with an event only if a time mention occurs in the same sentence as the event mention, or in an adjacent sentence with an overt temporal connective. This is a serious drawback because many events in unstructued text, including HUMINT, are not associated with explicit time information. A knowledge-rich solution entails using world knowledge resources with many levels of analysis (e.g., discourse parsing) in order to construct an Event/Time Network, which connects and partially orders all temporally related event mentions and time expressions in the document. This is unfeasible in the context of information extraction systems, which rely on relatively shallow structural analyses to rapidly process large volumes of text. As an alternative, in Phase I, we propose to develop a knowledge-light solution for automatically constructing the Event Network for an input document. By means of corpus-based empirical research, our goal is to identify surface features that may correlate with the presence and type of an event-event link. Such features may then be used in a machine learning approach or, where the correlation is strong enough, in the development of pattern-based solutions to the Event/Time Graph construction problem.