A large number of user reviews in the internet contains valuable information on services and products; for this reason, there is interest in automatically understanding such reviews. Sentiment Classification labels documents according to the feelings they express; instead of classifying a document into topics (sports, economics, etc), one attempts to tag the document according to overall feelings. Compared to the accuracy of traditional text categorization methods, sentiment classifiers have shown poor performance. We argue that such bad results are due to an improper representation of reviews. We describe a weakly supervisedmethod that converts raw text into an appropriate represen- tation, and show how techniques from information retrieval can acquire labeled data and process data using Markov logic. We report results on sentence classification and rating prediction that support our claims.