Have you considered using PLSA (Probabilistic latent semantic analysis)? Effectively, this method does very similiar thing as you aim to do - it chooses a set of words that represen a topic of a document.