Posterior calibration and exploratory analysis for natural language processing models