Recognizing Situation Patterns from Self-Contained Stories*

We propose extracting information about characters and actions from a self-contained story, such as news reports. This information is stored in structure patterns called situations. We show how these situation patterns can be constructed by unifying the constituents of sentence analysis with knowledge previously stored in Typed Feature Structures. These situations can be in turn used subsequently in the form of knowledge. The combination of situations constructs a supra-structure that represents the understanding of a factual report. The main pattern structure can be used to answer questions about facts and their participants.