Can prior knowledge help graph-based methods for keyword extraction?