FlorUniTo@TRAC-2: Retrofitting Word Embeddings on an Abusive Lexicon for Aggressive Language Detection