Constructing informative prior distributions from domain knowledge in text classification