Emoji are pictographs commonly used in microblogs as emotion markers, but they can also represent a much wider range of concepts. Additionally, they may occur in different positions within a message (e.g. a tweet), appear in sequences or act as word substitute. Emoji must be considered necessary elements in the analysis and processing of user generated content, since they can either provide fundamental syntactic information, emphasize what is already expressed in the text, or carry meaning that cannot be inferred from the words alone. We collected and annotated a corpus of 2475 tweets pairs with the aim of analyzing and then classifying emoji use with respect to redundancy. The best classification model achieved an F-score of 0.7. In this paper we shortly present the corpus, and we describe the classification experiments, explain the predictive features adopted, discuss the problematic aspects of our approach and suggest future improvements.