ToxiFusionNet: an explainable transformer-based multi-task framework for hate speech and fake narrative detection in Hindi–English code-mixed text

The classification of Hindi-English (Hinglish) code-mixed hate speech (HS) is greatly affected by the spread of fake news on social media. However, key challenges include the need for high-quality annotated code-mixed data and difficulties in handling evolving slang and cultural expressions. This work introduces ToxiFusionNet, a transformer-based model that can detect and analyse toxic, fake and hateful Hindi-English code-mixed posts on social media. Data were collected from online sources using application programming interfaces (APIs) and web-scraping tools. The data were then examined for duplicate or near-duplicate posts. Subsequently, to normalize Hinglish code-mixed content, the pipeline removes emojis, performs transliteration and translation, eliminates stop words, tokenizes the text, and applies stemming. The annotation process comprised two stages: first, determining whether the content was fake or hateful; and second, grouping the content according to the target (religion, caste, region, gender and cyberbullying) and the severity of hate (low, medium, high). The ToxiFusionNet integrates part-of-speech (POS) tags, sentiment scores, word-to-vector (Word2Vec) embeddings and Toxic Bidirectional Encoder Representations from Transformers (ToxicBERT) derived indicators to combine linguistic, semantic, and toxicity-related information. Its multi-task architecture incorporates task-specific loss functions, layer pruning, dropout, batch normalization, and explainable artificial intelligence (XAI) techniques to support model interpretation. On the internal test set, the framework achieved an accuracy of 0.980 for fake/non-fake classification and an F1-score of 0.978 for fake class. For hate/non-hate detection, it achieved a weighted F1-score of 0.975. The reported accuracy of 0.980 represents an improvement of 2.2 percentage points over HateBERT (0.958), confirming its effectiveness in detecting complex code-mixed HS. External evaluation on the HASOC 2021 dataset yielded an accuracy of 0.968 and F1-csore of 0.965. Human evaluation of token-level explanations yielded Cohen’s kappa values of 0.76, 0.79, and 0.81 for integrated gradients (IG), SHapley Additive exPlanations (SHAP), local interpretable model-agnostic explanations (LIME) respectively. These results suggest that ToxiFusionNet may support the identification and prioritization of potentially harmful code-mixed content for human review. Its suitability for near-real-time deployment requires further validation under operational under conditions.