Entity Profile Extraction from Large Corpora

Information Extraction (IE) has two anchor points: (i) entity-centric information leads to an Entity Profile (EP); (ii) action-centric information leads to an Event Scenario. Based on a pipelined architecture which involves both document-level IE and corpus-level IE, a multi-level modular approach to EP extraction from large corpora is described: (i) named entity tagging; (ii) three-level pattern matching for extracting the underlying correlated entity relationships; (iii) co-reference; (iv) document-internal merging of entity relationships into discourse EPs; and (v) cross-document fusion of EPs. The approach achieves around 90% precision and 50%-70% recall for major EP relationships. The significance of EP enhanced by cross-document fusion is demonstrated.