Using Stylistic Topic Models to Detect Deception Through Unusual Linguistic Activity

Analysts are faced with the challenge of sifting through enormous quantities of documents, blog posts, communications, etc. to find deceptive behaviors. We propose novel techniques for efficiently and automatically detecting deception on large data with high accuracy by using methodologies from both stylometry and topic modeling. This combined approach will learn models of authors and will detect unusual behavior based on their unconscious writing style or their topical content, or a combination of both. A comprehensive system will make the algorithmic results accessible through a web service to an intuitive user interface with search, drill-down, and cross-referencing with custom visualizations. This will allow analysts to quickly see the current big picture activity and also to discover particular events or trends of interest. The text analysis expertise of University of California Irvine and the software and visualization expertise of Kitware will provide the correct skill set to build these tools. Phase I will assess the feasibility of the algorithmic and visualization techniques needed for this system.