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

Relatively few event mentions in texts can be anchored to a precise time. The time-stamp of an event is thus an unknown, subject to constraints that we can extract from the text. These constraints can form a basis for inference over events and times mentioned in a document. We propose to develop a system to extract temporal constraints from text and collect them in a Temporal Constraint Store, from which we can derive new constraints by inference and constraint satisfaction techniques, test for inconsistency, and ultimately derive as refined an Event Network as possible, given information available in the text. We will leverage the resources available to us, especially the TimeBank corpus, to exhaustively survey and classify the various contexts in which event-event relationships are expressed in texts. Descriptions of these contexts will be formalized and implemented as recognizers for those contexts, from which we will derive temporal constraints to add to the store. Once the constraints in the constraint store have been reduced as far as possible, the resulting distilled information will be added to the World Model constructed by our Information Extraction engine. From there it can be used to drive visualization applications and other intelligence analysis tools.