Enabling Visualization of Event Information from Unstructured Text

The value in the ability to visualize information about events extracted from unstructured text rests on high-precision, high-recall event extraction technology. Event extraction must provide information about properties of the events themselves (times, locations, and participants) as well as information about connections and similarity between events. The current state of the art in event extraction is the identification and keyword-based grouping of verb-anchored event mentions. The work proposed here goes beyond this to pursue the extension of event extraction to (1) detection and linking of nominal event mentions and (2) data-driven automatic clustering of events by type. Detecting nominal event mentions can increase recall of event detection systems, in particular of the most important events in a document and time-stamped event mentions, and will allow the application of coreference resolution techniques developed for entity tracking to events. Unsupervised and weakly supervised event clustering techniques will obviate the need for creating and maintaining a single, static ontology of event types. It will also allow users to define event types by example and search for them on-the-fly. The guiding principle in this effort is the desire for solutions that will add value to event visualization for the end analysts.