Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{198994,
        author = {PROF.DHANANJAY R RAUT and Madhur Vinod Shinde and Vishal RajendraPrasad Yadav and Aman Suresh Singh and Harsh Jayendra Sakpal},
        title = {Implementation of a Real-Time, Multilingual, Emotion-Aware Cyberbullying Detection System Using Multi-Teacher Knowledge Distillation and Explainable AI},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {10530-10544},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=198994},
        abstract = {The rapid expansion of social media in India has led to an explosion of code-mixed toxicity, presenting a "Trinity of Challenges" for automated moderation: deep transformer models are too computationally expensive for real-time streams, lightweight models lack the contextual awareness to detect sarcasm or emotional nuance, and traditional classifiers act as opaque "black boxes."
To address this, we implement a highly optimized, multilingual cyberbullying detection system utilizing Multi-Teacher Knowledge Distillation (MTKD). Our approach compresses an ensemble of heavy transformers (mBERT, XLM-R, and MuRIL) into a 0.91 MB XGBoost Student model utilizing a 21,384-dimensional feature space.
To regain contextual intelligence, the student model is augmented by a 4-Rule Hybrid Fusion Engine that dynamically calibrates threat levels using auxiliary BiGRU (Sarcasm) and XLM-R (Emotion) networks. Finally, to ensure operational transparency without sacrificing processing speed, we introduce a novel 5-Stage Short-Circuit Explainable AI (XAI) pipeline that bypasses computationally expensive SHAP calculations using O (1) dictionary heuristics.
Evaluated on a 3,272-sample multilingual social media dataset spanning 14 languages, our system achieves sub-second inference latency (~300ms) and massive model compression (~1000x) while achieving a robust F1-score of 0.93, proving its efficacy for enterprise edge-device deployment.},
        keywords = {Cyberbullying Detection, Multi-Teacher Knowledge Distillation (MTKD), Multilingual NLP, Explainable AI (SHAP), Affective Computing, Code-Mixing.},
        month = {April},
        }