Classi cation of Corporate and Public Text

In this project we try to tackle the problem of classifying a body of text as a corporate message (text usually sent to a private, select audience, usually sent in a company setting and often related to company information) and public messages (text that is more open and can be broadcasted to a larger audience). Corporate messages usually contain sensitive, private data that would put an organization or individual at risk if leaked. This includes messages about nancial trades, company deals, and business meetings. Public messages, however, are less secretive and more casual by nature. This problem is very similar to Data Loss Prevention, a security issue that involves systems that identify, monitor and protect con dential data from leakage. While information about these systems is also con dential, the general industry techniques involve regular expressions, keywords, and hashing. Regular expressions are used to match data such as social security numbers, telephone numbers, and addresses. Keyword matching is used to identify a few words that are marked as private. And hashing works by hashing the substrings of private documents and classi es a new document as private if it contains a substring with a matching hash. Our problem is similar to DLP, but given our data set it would be erroneous to consider it as DLP. Instead of looking for con dential information, we look to see whether it would be in a corporate message. Still, this work can help shed light on DLP, perhaps by improving the keywords used in DLP techniques. In or comparison of text classi iers, we used Naive Bayes, Logistic Regression, and Support Vector Machine classi ers and found that SVMs showed consistently better results. However, noticing that the corporate and public messages were centered around certain topics, we used LDA to improve our logistic regression model, which has probablistic inclinations. While we found that with LDA logistic regression results improved, they were still slightly below SVMs.