Cabil: Lightweight Imbalance-Aware Multilingual Hate Speech Detection

Hate speech on social media threatens both individual well-being and social cohesion, yet automated detection remains difficult because hateful content is context-dependent, often code-mixed or multilingual, and appears far less frequently than offensive or neutral posts. This paper proposes CABiL, a lightweight hybrid deep learning framework that combines a Convolutional Neural Network (CNN), a Bidirectional Long Short-Term Memory (BiLSTM) network, and a custom Bahdanau-style self-attention mechanism for hate speech classification. Pre-trained GloVe embeddings provide the semantic input representation; the CNN layer extracts local n-gram patterns indicative of hateful or offensive language; the BiLSTM layer models bidirectional long-range dependencies; and the attention layer adaptively re-weights token representations before a softmax classifier assigns the final label. To address the severe class imbalance characteristic of hate speech corpora — hate speech makes up only about 1.16% of the primary Kaggle benchmark — the framework incorporates cost-sensitive class weighting and SMOTE-based augmentation, which together raise hate-class F1 from 0.424 to 0.527 and macro-F1 from 0.728 to 0.766. On the Kaggle Hate Speech and Offensive Language benchmark, CABiL attains 90.06% accuracy and a weighted F1-score of 0.90, outperforming CNN, LSTM, CNN–LSTM, and CNN–BiLSTM baselines, and matching BERT-based models with roughly 56× fewer parameters (1.97M vs. 110M) and about 5.6× lower per-example inference latency. Cross-dataset evaluation on ETHOS (multi-label English), HASOC-Dravidian (code-mixed Hindi/Tamil/Malayalam–English), and LAHM (six-language multilingual corpus) confirms that CABiL generalizes across linguistic settings, achieving 87–92% accuracy and 0.79–0.90 macro-F1. These results indicate that a compact hybrid CNN–Attention–BiLSTM architecture, combined with imbalance-aware training, offers an efficient and robust alternative to heavyweight transformer models for real-time, resource-constrained hate speech moderation.