Story Creation and Inference through Bayesian Extraction (SCRIBE)

Without a deep understanding of causes, context, and meaning of the events they observe, Warfighters are unable to reliably develop or execute effective courses of action. Unfortunately, traditional methods of accessing this information require significant effort searching large volumes of unstructured text from varied open sources. Advanced methods for coping with this problem face several challenges: 1) Existing event and entity extraction technology is constrained by expensive language-specific models and predefined event ontologies. 2) Existing methods of organizing information do not respect causal relationships or source quality. 3) There are no effective, general-purpose methods of organizing, visualizing, and navigating a complex web of causally linked events. We propose a system called Story CReation and Inference through Bayesian Extraction (SCRIBE). SCRIBE"s explanatory cause-and-effect models are developed through three stages. The event discovery stage uses an innovative unsupervised method which can operate on the spectrum of language genres, from news articles, to micro-blogs, to conversation transcripts in any language. The event linking phase uses statistical cues to link events through the evolving fingerprint of entity and topic references woven into storylines of real-world activities. Finally, SCRIBE"s visualization step provides users with interactive visualizations offering views of underlying event sequences at varying granularities.