Multilingual Indic Cyberbullying Detection with Text-Meme Fusion, Severity Grading, LIME Explainability, and RealTime Streamlit Deployment
International Journal of Drug Delivery Technology · 27.06.2026 · 16 (58s)
DOI: 10.25258/ijddt.16.58s.22 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 27.06.2026. Verfasst von Nishikant Vetal und 1 weiteren Autor:innen.
Abstract (Originalsprache)
A growing number of users face online harassment through India's diverse digital spaces, where communication blends Hindi, Hinglish, Tamil, Telugu, Marathi, alongside many regional tongues. While most tools target English content, they fail to catch harmful expressions in native scripts, creating blind spots for vast communities. In the course of an M.Tech initiative based in Pune, a functional model emerged - one capable of identifying abusive behavior across five prominent Indic languages. Performance varied: accuracy peaked at 87.4% for Hindi, dipped slightly for Marathi (81.9%) and Hinglish (80.1%), then lowered further for Telugu (67.7%) and Tamil (65.0%), averaging 74.6%. Validation unfolded via live examination - a Streamlit-powered interface processed genuine hostile posts, including inflammatory tweets and imagebased memes The system does much more than basic classification. It uses OCR to pull text from meme images (59.0% accuracy on exam stress memes with 34 OCR regions), runs an IndicBERT ensemble of 4 fine-tuned transformer models, grades severity as Severe (≥85%), Moderate (65-85%), or Safe, and provides LIME explanations showing exactly which toxic words matter most - like "bastard[+0.82]" in Marathi or "इिडयट[+0.65]" in Hindi. My deployed Streamlit dashboard handles live single predictions, bulk CSV analysis of 47k tweets, and generates human-readable explanations in under 300ms - ready for real platforms right now. Looking at four dozen studies, more than half concentrate solely on spotting English text. A small fraction - only one in ten - addresses memes mixing images and words. Just three out of twenty tackle how severe the bullying is. None bring together support for Indian languages, combining visuals and text, showing why decisions are made through LIME maps, along with real-world usability. What stands apart here is a working model built for scale, filling every missing piece. Built for India's vast online population, it delivers clarity across many native tongues, shows its reasoning clearly, runs reliably, and marks the first full solution of its kind.
Zitieren
APA
Vetal, N., & Vasgi, B. (2026). Multilingual Indic Cyberbullying Detection with Text-Meme Fusion, Severity Grading, LIME Explainability, and RealTime Streamlit Deployment. International Journal of Drug Delivery Technology, 16(58s). https://doi.org/10.25258/ijddt.16.58s.22
Vancouver
Vetal N, Vasgi B. Multilingual Indic Cyberbullying Detection with Text-Meme Fusion, Severity Grading, LIME Explainability, and RealTime Streamlit Deployment. International Journal of Drug Delivery Technology. 2026;16(58s). doi:10.25258/ijddt.16.58s.22
BibTeX
@article{Vetal2026,
author = {Vetal, Nishikant and Vasgi, Bharati},
title = {Multilingual Indic Cyberbullying Detection with Text-Meme Fusion, Severity Grading, LIME Explainability, and RealTime Streamlit Deployment},
journal = {International Journal of Drug Delivery Technology},
year = {2026},
volume = {16},
number = {58s},
doi = {10.25258/ijddt.16.58s.22}
}
EUVIMED