Cross-Dataset Evaluation of mBERT and MuRIL for Hinglish and Hindi-English Harmful Speech Detection

This study compares mBERT and MuRIL for binary harmful-speech classification across three Hindi-English/Hinglish dataset settings: a Kaggle Hinglish hate subset, a code-mixed offensive-language dataset, and THAR targeted religious hate. Matched multi-seed experiments show that model performance varies substantially across datasets: mBERT performs better on some settings, while MuRIL performs better on THAR. Cross-dataset evaluation, TF-IDF baselines, and mixed-training experiments suggest that label definitions, platform, topic, and script mix materially affect model behavior. The results therefore do not support a universal claim that either mBERT or MuRIL is superior for Hinglish harmful-speech detection; instead, they highlight the importance of dataset choice and cross-dataset robustness when evaluating multilingual harmful-speech classifiers.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-16
DOI
https://doi.org/10.5281/zenodo.24285266
Primary Topic
Hate Speech and Cyberbullying Detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Cross-Dataset Evaluation of mBERT and MuRIL for Hinglish and Hindi-English Harmful Speech Detection

Om Patnaik
1 links
Zenodo (CERN European Organization for Nuclear Research)
Hate Speech and Cyberbullying Detection
preprint

Cross-Dataset Evaluation of mBERT and MuRIL for Hinglish and Hindi-English Harmful Speech Detection

Om Patnaik
preprint en

Abstract

This study compares mBERT and MuRIL for binary harmful-speech classification across three Hindi-English/Hinglish dataset settings: a Kaggle Hinglish hate subset, a code-mixed offensive-language dataset, and THAR targeted religious hate. Matched multi-seed experiments show that model performance varies substantially across datasets: mBERT performs better on some settings, while MuRIL performs better on THAR. Cross-dataset evaluation, TF-IDF baselines, and mixed-training experiments suggest that label definitions, platform, topic, and script mix materially affect model behavior. The results therefore do not support a universal claim that either mBERT or MuRIL is superior for Hinglish harmful-speech detection; instead, they highlight the importance of dataset choice and cross-dataset robustness when evaluating multilingual harmful-speech classifiers.

Zenodo (CERN European Organization for Nuclear Research)
Hate Speech and Cyberbullying Detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon