Techniques are described for generating a monosemous (i.e., single sense) keyword list associated with a particular domain (e.g., a medical or financial domain) for document classification. An input term frequency dictionary, a candidate keyword list, and a document corpus may be used to generate the keyword list. A collection of documents is divided into two sets, one related to a target domain and one not. A statistical approach may be used to evaluate each term in the candidate list to determine a measure of how monosemous each remaining candidate term is, i.e., how strongly the term (or short phrase) identifies with a single sense. Terms with a primarily single sense related to the target domain are added to the monosemous keyword list. The keyword list may be used to identify documents associated with the domain, allowing, the appropriate protections to be applied to the document (e.g., do not send outside an enterprise boundary or permit copying).