Rapid visualization of events existing in unstructured data requires a system with the ability to accurately detect events and their arguments in such data. Currently, many systems that perform this task do so by relying on only sentential context. Those systems that do rely on document-wide context are more accurate but suffer in terms of efficiency in that they need to process the input document repeatedly, or that in order to process the current input document they require processing of related documents. We experiment with novel approaches for event extraction that rely on document-wide context. Here, event extraction includes event mention detection, event argument detection, and event coreference resolution. One approach is maximum entropy modeling with document-wide features. Another approach models the document as a Dynamic Bayesian Network. They offer the promise of event extraction with higher recall, precision, and speed than previous systems.
Benefit: The main anticipated benefit of this work includes the development of a system that (i) extracts event mentions and event arguments from unstructured data with higher recall and precision, and (ii) performs event coreference resolution on these event mentions with higher accuracy than was previously attainable. Furthermore, the document-wide classification techniques developed on this project might be transferred to improve the accuracy of other information extraction modules.
Keywords: Event Extraction, Event Coreference, Location Extraction, Information Extraction, Text Extraction, Spatial Annotation
