Improving NMF clustering by leveraging contextual relationships among words