Assessing users' interactions for clustering web documents: a pragmatic approach