A comparative study of machine learning models for hate speech and stereotype detection in Italian texts

This study presents a comparative analysis of various machine learning models for hate speech and stereotype detection in Italian texts. The research utilises datasets from the HaSpeeDe task proposed by EVALITA in 2020. Multiple text representation techniques are evaluated, including non-lexical linguistic information, bag of words, n-grams (characters, words, and part-of-speech tags), word embeddings, and a neural language model (BERT). The study compares the performance of these models in different metrics such as accuracy, precision, recall, and F1-score. The results indicate that character n-grams and the neural language model (BERT) generally outperform other techniques, with BERT achieving the highest accuracy (76%) for the detection of hate speech and character n-grams performing the best for the detection of stereotypes (72% accuracy). The research highlights the challenges in detecting stereotypes compared to hate speech and emphasises the importance of context in classification tasks.