exploiting headword dependency and predictive clustering for published presentations and documents on DocSlides.