UR NLP @ HaSpeeDe 2 at EVALITA 2020: Towards Robust Hate Speech Detection with Contextual Embeddings

We describe our approach to addressTask A of the EVALITA 2020 Hate SpeechDetection (HaSpeeDe2) challenge.Wesubmitted two runs that are both based oncontextual embeddings – which we hadchosen due to their effectiveness in solvinga wide range of NLP problems. For ourbaseline run we use stacked embeddingsthat serve as features in a linear SVM. Oursecond run is a simple ensemble approachof three SVMs with majority voting. Bothapproaches outperform the official base-lines by a large margin, and the ensembleclassifier in particular demonstrates robustperformance on different types of test datacoming 6th (out of 27 runs) for news head-lines and 10th (out of 27) for Twitter feeds.