Lightening the load of document smoothing for better language modeling retrieval