Exploring statistical properties for semantic annotation: sparse distributed and convergent assumptions for keywords