Enabling Visualization of Event Information from Unstructured Text

We propose to address two outstanding problems encountered in event extraction and visualization. The first focuses on recognizing events in nominal expressions. This is important because nominal coreference among event mentions offers a way of propagating information (e.g. time). It is difficult because not all ``event words'' denote actually occurring events, or even events at all (e.g., `devastation' may refer to the result of an event). The second problem is linking participants to events. When participant roles are defined at too low a level, information extraction systems experience a knowledge bottleneck making them expensive to customize with new event types. We believe that the development of the FrameNet database offers solutions here since the ``frame'' level of detail is more general than event specific template slots. A frame-based approach also simplifies customization: new events are defined mainly by example. Both problems (recognizing events and linking event participants) will be addressed using unsupervised machine learning techniques, mainly bootstrapping techniques. These techniques learn extraction patterns by using a high precision, low recall classifier to automatically construct a very large training corpus. This research if successful will significantly enhance the ability to identify events that can be displayed on a timeline.