PROXIMIX: A PROTOTYPE-DRIVEN PROXIMITY-AWARE META-EMBEDDING FRAMEWORK FOR FAUX-HATE DETECTION

Abstract

Faux-Hate detection in code-mixed Hinglish text is a complex multilingual classifi cation task due to the noisy social media language, transliteration variation, semantic ambiguity and the large class imbalance. In this work, we introduce a prototype-based proximity-aware meta-embedding framework for multi-class fake news and hate speech detection in low-resource code-mixed scenarios, dubbed ProxiMix. The proposed model is a fine-tuning of XLM-RoBERTa’s encoder to learn powerful cross-lingual contextual representations, followed by a learnable prototype bank constructed using K-Means clus tering on the auxiliary corpora. Such prototypes represent semantic regions of the class and enable the model to learn more complex decision boundaries than regular fixed em beddings. ProxiMix introduces cosine similarity, multi-head cross-attention, and feature fusion, which enables the representation learning of prototype-level proximity informa tion to be fused with the contextual CLS embedding to further refine the representation learning. The framework also tackles imbalance and optimization instability by applying focal loss, label smoothing, prototype regularization, Monte Carlo dropout and isotonic calibration in five-fold cross-validation. Experimental results conducted on the Faux-Hate corpus show the mean Macro F1 is 87.00% for fake detection and 87.61% for hate detec tion with a standard deviation of 0.0064 for the folds, demonstrating good consistency and robustness. Additionally, the performance difference between the two tasks is smaller in the model, suggesting that the prototype-aware semantic fusion approach leads to better generalization under low-resource and noisy multilingual NLP tasks.