DH-FBK @ HaSpeeDe2: Italian Hate Speech Detection via Self-Training and Oversampling

We describe in this paper the system submitted by the DH-FBK team to the HaSpeeDe evaluation task, and dealing with Italian hate speech detection (Task A). While we adopt a standard approach for fine-tuning AlBERTo, the Italian BERT model trained on tweets, we propose to improve the final classification performance by two additional steps, i.e. self-training and oversampling. Indeed, we extend the initial training data with additional silver data, carefully sampled from domain-specific tweets and obtained after first training our system only with the task training data. Then, we retrain the classifier by merging silver and task training data but oversampling the latter, so that the obtained model is more robust to possible inconsistencies in the silver data. With this configuration, we obtain a macro-averaged F1 of 0.753 on tweets, and 0.702 on news headlines.