Automating the Application of Deception Detection Heuristics to Unstructured Data

We propose to construct a deception detection system which will exploit scaffolding provided by a collection of largely domain-independent deception detection heuristics. These heuristics, integrated through a novel evidential reasoning system, will provide the proposed system, called Skeptic, with a significant advantage over purely inductive methods by allowing it to exploit the adversarial nature of the problem. Whereas previous systems have only provided coarse level judgments regarding the deceptive text communications, Skeptic will employ a mix of lightweight natural language processing and information extraction techniques to allow for the detection of misleading information present in otherwise truthful communications. Further, Skeptic will adapt over time, which means it can be deployed early, and mature as the understanding of the different operational contexts matures. In this work we will exploit our team’s substantial software assets and experience in the areas of text analysis and machine learning, as well as our very-specifically related experience developing a system for detecting a particular class of deception called “stock pumping.” Given this foundation we are able to propose an aggressive work plan that will result in a proof-of-concept demonstration against multiple existing datasets.