Approximate learning of parsimonious Bayesian context trees