POPSTAR at RepLab 2013: Polarity for Reputation Classification

Abstract. This paper describes our participation in the Polarity for Reputation classification task of RepLab 2013. Our system leveraged on a set of components previously developed for a Twitter message polarity classifier. Following a super-vised approach, a Logistic Regression classifier is trained from annotated data. A refined language model is used to represent tweets in terms of a vocabulary consisting only of the most informative terms with word features weighted using a measure from the Information Retrieval field. To help reduce the sparseness of the feature vector, the model is enriched with another, more compact, represen-tation of the words. Finally, we extract features to capture the use of informal and affective language. Our approach ranked in the top three for all the metrics, showing that the strategies for Twitter Sentiment Analysis are useful for the task of Polarity for Reputation classification.