Unsupervised Learning of Interpretable Dialog Models