Improving a textual deception detection model

In intelligence, law enforcement, and, increasingly, organizational settings there is interest in detecting deception; for example, in intercepted phone calls, emails, and web sites. Humans are not naturally good at detecting deception, but recent work has shown that deception is actually readily detectable - using markers that humans don't see but which software can readily compute. Pennebaker's model suggests that deceptive communication is characterized by changes in the frequency of four kinds of words: first-person pronouns, exception words, negative emotion words, and action words.We investigate what can be learned about the deception model by applying it to a large corpus of Enron emails. We show that each of the four kinds of words in the Pennebaker model acts as a separate latent factor for deception, rather than having their effects mixed together.