perhaps you can integrate a part-of-speech tagger from nltk or montylingua to help it pick the right kind of word.

http://www.nltk.org/ http://web.media.mit.edu/~hugo/montylingua/