I'd be curious how the output of your approach compares to merely classifying based on what keywords are contained in the text (given that AFAICT you're simply categorizing rather than trying to extract precise meaning).