Thumbnail: click to enlarge

ONSITE System Architecture

KBSI has been awarded funding from the Defense Advanced Research Projects Agency (DARPA) to research, design, and demonstrate enabling technology for Open-Source Information Tactical Exploitation (ONSITE).  The ONSITE initiative, in Phase II, applies an innovative approach to natural language processing (NLP) that aims at achieving state of the art processing rates for the understanding of natural language in support of military tactical operations.  DARPA's goal is to improve natural language processing speed and efficiency despite constrained computational resources, accelerated operational timelines, and specific intelligence objectives.  Improving the speed and efficiency of NLP allows war fighters to more quickly process data in the bid for tactical advantage.

The ONSITE technology addresses the shortcomings of current methods for processing open-source, natural-language information.  While the automated support of intelligence analysis has matured with the emergence of large-scale storage and data-processing capabilities, and sophisticated natural-language processing algorithms, tactical applications making use of NLP are hampered by the processing speed of current algorithms and the restrictions of tactical computing platforms.  KBSI's approach deviated from traditional statistical and grammar rule-based approaches to shallow NLP (e.g., from part-of-speech and phrase chunking) by instead using a specially constructed ontology that led to higher processing throughput and increased semantic resolution of extracted information. 

The initial phase of the ONSITE initiative investigated and developed the following:

	optimized heuristics that limit the amount of math-related processing (especially floating-point operations), leading to higher CPU utilization;
	heuristics that support the structuring of calculations to allow for better use of higher-speed CPU cache memory; and
	heuristics that improve so-called “shallow” natural language processing by requiring less system computation and increasing robustness despite noisy data.

In the intial phase, the semantically rich extractions that were generated focused on supporting influence operations that targeted enemy networks, both physical (existing in the real world) and virtual (existing online).