Detecting hostile posts in Hindi on social media is challenging due to linguistic variability, informal usage, and code-switching in the Devanagari script. While prior efforts have addressed hostility detection, few have focused on fine-grained, multi-label classification using ensemble strategies suited for Hindi. In this work, we propose a novel ensemble-based framework for hostility classification across four categories: Defamation, Fake, Hate, and Offensive. The system integrates Bagging, Boosting, Simple Majority Voting, and Weighted Majority Voting over contextual embeddings derived from MuRIL, IndicBERT, and HindiBERT models. Principal Component Analysis (PCA) reduces dimensionality and computational complexity. Evaluation on the CONSTRAINT-2021 dataset demonstrates that our model achieves F1-scores of 0.8874 (Defamation), 0.9532 (Fake), 0.8653 (Hate), and 0.8790 (Offensive), outperforming prior work and recent benchmarks. The proposed model shows relative improvements of 45%, 13%, 28%, and 25% across the respective hostility classes. This demonstrates the effectiveness of combining multilingual transformer embeddings with ensemble strategies for hostile content classification. The approach offers a scalable, language-sensitive solution for detecting hostility in Hindi social media, supporting more respectful and safer digital interactions.
