The project aims at a computational system for detecting unintentional inferences in casual unsolicited and unrestricted verbal output of individuals, potentially responsible for leaked classified information to people with unauthorized access.  We are interested in automatic extraction of hidden semantic information from the casual and unsolicited verbal output of a “person of interest” (POI), both written (blogs, Facebook, Twitter, etc.) and oral (taped conversations), over any period of time. 

We claim that -- and make use of it -- any information in a text is either just additional (previously unknown) information, or it overwrites the existing (salient) information that comes from some background knowledge and serves as a default. In other words, any statement has its purpose, and these two are most frequent and significant. The communicational choice between overwriting and simply adding information depends on the knowledge of what information is implicitly salient for the speaker or whether enough priming has been achieved by the explicitly communicated text, respectively. We are most interested in the cases of the absence of information when it seems that it should be there because this may indicate a default for the speaker. The absence of information is determined through the need of additions or overwrites in a message relative to the knowledge of the people communicating.