This paper describes our submission for SemEval2013 Task 2: Sentiment Analysis in Twitter. For the limited data condition we use a lexicon-based model. The model uses an affective lexicon automaticallygeneratedfrom a very large corpus of raw web data. Statistics are calculated over the word and bigram affective ratings and used as features of a Naive Bayes tree model. For the unconstrained data scenario we combine the lexicon-based model with a classifier built on maximum entropy language models and trained on a large external dataset. The two models are fused at the posterior level to produce a final output. The approach proved successful, reaching rankings of 9th and 4th in the twitter sentiment analysis constrained and unconstrained scenario respectively, despite using only lexical features.