Textually retrieved event analysis toolset

Of interest to the Intelligence (INTEL) community in its focus of the Global war on Terror (GWOT), is a capability to assist reasoning about events in the context of emerging patterns of behavior. An event ontology that decomposes event interactions into (actor, action, target) triplets is proposed for this purpose. Within this ontology battlespace objects (actors, targets) and their interrelationships (including actions) are formulated as graphs. The vertices represent the building blocks comprised of the individuals or capital assets necessary for war-fighting. The links or edges represent relationships between the coupled nodes that identify a specific type of association. Different links within a graph may represent different types of associations2. The increasing importance of human intelligence (HUMINT) and open source intelligence (OSINT) in combating asymmetric threats requires a capability to extract and reason about events described in free text. Attribute-based entity correlation - Level I Data Fusion (Hall and Llinas, 2001) - and inexact pattern matching techniques (e.g., subgraph isomorphism) may be sufficient to reason about the actor and target components of an event interaction triplet. Processing event description extracted from free (unstructured) text, however, requires semantic normalization and correlation to discern commonality among reported activities. A semantic comparator function is proposed to measure the similarity between the action components of extracted event triplets. The paper reviews existing work in this area such as Jiang and Conrath (Jiang and Conrath, 1997), who combine a lexical taxonomy structure (such as Princeton's WordNet) with corpus statistical information to quantify the semantic, distance between concepts. The Jiang and Conrath measure combines an edge-based approach of the edge counting scheme with a node-based approach based upon information content calculation. This paper investigates the feasibility of replacing a distributional analysis of corpus data with the Lavalette distribution (a Zipfian distribution). A prototype free text, event analysis system was developed to process simulated human intelligence information reports based on this investigation. The system was able to identify event relationships from semantic comparisons of extracted and decomposed events.