Sparse, Contextually Informed Models for Irony Detection: Exploiting User Communities, Entities and Sentiment