Sentiment prediction from Twitter is of the utmost interest for research and commercial organizations. Systems are usually using lexicons, where each word is positive or negative. However, word lexicons suffer from ambiguities at a contextual level: the word dark is positive in dark chocolate and negative in dark soul, the word lost is positive with weight and so on. We introduce a method which helps to identify frequent contexts in which a word switches polarity, and to reveal which words often appear in both positive and negative contexts. We show that our method matches human perception of polarity and demonstrate improvements in automated sentiment classification. Our method also helps to assess the suitability to use an existing lexicon to a new platform (e.g. Twitter).