It's a very limited task: take a document and classify it into one of (for example) 10 or so categories. Things like detecting certain words can do pretty well in some cases. Things that compress well have the occurrence of common substrings.