... sN)}. The other variables, y, z, and a, are latent. For all models, we estimate param-eters to maximize the marginal likelihood of the ob-served reviews. For example, for RSA-MixHMM,we estimate ... Wecreated a new recipe data set, and manually labeled a random sample to evaluate our model and severalbaselines. We obtain 90.1% F1 for predicting refine-ments at the review level, and 77.0% F1 for ... StanfordParser4. Based on a quick manual inspection, do-main shift and ungrammatical sentences do cause a significant degradation in parsing accuracy whencompared to in-domain data. However,...