Hate Speech is not Free Speech: Explainable Machine Learning for Hate Speech Detection in Code-Mixed Languages

The increase in connectivity provided by social media platforms comes with several disadvantages. It has become surprisingly easy for ill-intentioned individuals to stalk, harass and threaten others online on the basis of race, gender, religion, etc. Artificial intelligence models provide a valuable solution to the problem by automatically filtering such content. However, research for hate speech detection in low-resource languages is still nascent. To promote research in the area, several shared tasks are held in various languages and attract contributions with novel approaches. In this study, we use the English and Hindi code-mixed datasets provided by the HASOC Identification of Conversational Hate-Speech in Code-Mixed Languages (ICHCL) shared task to train and evaluate machine learning models and compare the performance of different vectorization techniques and model parameters. The Explainable Artificial Intelligence (XAI) technique Local Interpretable Model Agnostic Explanation (LIME) is also applied to the model results to ascertain if the model is behaving in a coherent and logical manner. The highest accuracy achieved is 67.61 % using Bernoulli Naive Bayes with count vectorizer and TF-IDF. The analysis suggests that many of the models may be heavily influenced by the presence of slur words while classifying a statement as hateful and offensive.