Detecting Deception in On and Off-line Communications

As online services become increasingly prevalent, detection of deception in electronic communication becomes vital to maintaining the many communities built around mutual trust between strangers (e.g. OkCupid, Quora, Yelp). Automating the detection process via machine learning has shown promise in recent years. This paper presents the results of using support vector machines (SVMs) and probabilistic context-free grammars (PCFGs) to classify text as truthful or deceptive. Experiments are conducted on two wellknown datasets of truthful and detective texts. A novel corpus is also developed via an online interrogation game, yielding additional insights regarding detecting deception in online conversations. The results of these experiments suggest that SVMs are superior to PCFGs at classifying deceptive text, and that the classification of deceptive text is highly dependant upon domain.