Concept-Based Event Extraction Utilizing Rich Semantics (CONVERSE)

This effort extends a novel approach to concept-based event extraction that leverages a rich substrate of semantic and conceptual knowledge in order to extract all of the essential information associated with events in text. In this work, we combine semantic information from a number of state-of-the-art text processing systems – including (1) word sense disambiguation systems, (2) semantic parsers (based on PropBank, NomBank, and FrameNet annotations), (3) within-document and (4) cross-document coreference resolution systems, (5) named entity recognition systems, and (6) discourse parsers -- in order to produce robust conceptual representations of events in any domain. These forms of knowledge are then used in conjunction with an active learning-based framework for open domain event extraction which can be rapidly customized to meet the particular information needs of a user. We plan to extend our Phase I work by incorporating new (1) statistical models for estimating the correctness of extracted information, (2) kernel-based methods for extracting the essential relations associated with event types, (3) unsupervised methods for named entity recognition, (4) inference-based methods for recognizing coherence relations, and (5) event coreference and event merging techniques for performing inter-sentential event extraction.