Indo-HateSpeech Analysis: A Multi-Level Hate Speech Classification Framework Using BERT Features and Machine Learning Models

The rise of social media has enabled individuals to express opinions widely, but it has also facilitated the spread of hate speech, posing a significant societal challenge. Effective automated hate speech detection systems are crucial to mitigating this issue. Addressing this issue in the context of diverse languages like those in India, this paper presented Indo-HateSpeech, a multi-level framework that combines contextualized embeddings from Bidirectional Encoder Representations from Transformers (BERT) with classical machine learning models for hate speech classification. The comprehensive dataset including emojis, emoticons, hashtags, and slang is necessary to identify hate speech on social media based on current trends. Therefore, the Indo-HateSpeech dataset is developed and collected from various posts on Instagram. This dataset contains hate speech sentences in English and Hindi in Indian Context for classifying hate speech into three categories: No Hate (HS0), Hate (HS1), and Extreme Hate (HSN). Contextual embeddings derived from the pre-trained BERT model to extract high-quality text features from text data with sentiment features. These embeddings are subsequently fed into traditional machine learning classifiers, including Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM), for classification. Experimental results demonstrate the effectiveness of BERT embeddings in enhancing classification performance in terms of accuracy, F1-score, and recall over traditional approaches, highlighting the framework's potential for scalable deployment in real-world applications