An important property of NMF is that it can produce a sparse representation of the data; however, in some applications, especially in text clustering, the sparse representation always consists of separated words, which cannot explicitly express the meaning of the basis vector.