# An Explainable System for Offensive Content Detection in Nastaliq Urdu: Corpus Curation, Transformer Benchmarking, and Word-Level Attribution

arXiv: https://arxiv.org/abs/2610.15795 · HTML: https://mlsanity.com/paper/2610.15795/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 22 of its 66 references that are in the map.
- First posted on arXiv 11 Oct 2026.

## Builds on (most cited first, 22 of 22)

- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://mlsanity.com/paper/1810.04805/) (2018)
- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [A Unified Approach to Interpreting Model Predictions](https://mlsanity.com/paper/1705.07874/) (2017)
- ["Why Should I Trust You?": Explaining the Predictions of Any Classifier](https://mlsanity.com/paper/1602.04938/) (2016)
- [XTREME: A Massively Multilingual Multi-task Benchmark for Evaluating Cross-lingual Generalization](https://mlsanity.com/paper/2003.11080/) (2020)
- [This Looks Like That: Deep Learning for Interpretable Image Recognition](https://mlsanity.com/paper/1806.10574/) (2018)
- [Automated Hate Speech Detection and the Problem of Offensive Language](https://mlsanity.com/paper/1703.04009/) (2017)
- [Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges](https://mlsanity.com/paper/2103.11251/) (2021)
- [Deep k-Nearest Neighbors: Towards Confident, Interpretable and Robust Deep Learning](https://mlsanity.com/paper/1803.04765/) (2018)
- [Interpretable Convolutional Neural Networks](https://mlsanity.com/paper/1710.00935/) (2017)
- [Offensive Language Identification in Greek](https://mlsanity.com/paper/2003.07459/) (2020)
- [Offensive Language and Hate Speech Detection for Danish](https://mlsanity.com/paper/1908.04531/) (2019)
- [TwHIN-BERT: A Socially-Enriched Pre-trained Language Model for Multilingual Tweet Representations at Twitter](https://mlsanity.com/paper/2209.07562/) (2022)
- [Arabic Offensive Language on Twitter: Analysis and Experiments](https://mlsanity.com/paper/2004.02192/) (2020)
- [A systematic review of Hate Speech automatic detection using Natural Language Processing](https://mlsanity.com/paper/2106.00742/) (2021)
- [Simple rules for complex decisions](https://mlsanity.com/paper/1702.04690/) (2017)
- [Balancing the Tradeoff Between Clustering Value and Interpretability](https://mlsanity.com/paper/1912.07820/) (2019)
- [Greybox XAI: a Neural-Symbolic learning framework to produce interpretable predictions for image classification](https://mlsanity.com/paper/2209.14974/) (2022)
- [AcME -- Accelerated Model-agnostic Explanations: Fast Whitening of the Machine-Learning Black Box](https://mlsanity.com/paper/2112.12635/) (2021)
- [HateCOT: An Explanation-Enhanced Dataset for Generalizable Offensive Speech Detection via Large Language Models](https://mlsanity.com/paper/2403.11456/) (2024)
- [Chinese Offensive Language Detection:Current Status and Future Directions](https://mlsanity.com/paper/2403.18314/) (2024)
- [Abusive and Threatening Language Detection in Urdu using Supervised Machine Learning and Feature Combinations](https://mlsanity.com/paper/2204.03062/) (2022)

## Abstract (arXiv)

Automated moderation systems are deployed at scale, yet they remain opaque to the users and moderators affected by their decisions, and their coverage is almost entirely confined to English. For Urdu written in the Nastaliq script, no such system exists, and the annotated resources it would be built on are drawn largely from Twitter/X, where character limits constrain how users write. Interpretability has received little attention, leaving it unclear what such models actually learn. This study addresses these gaps in three steps. First, we curate and publicly release a corpus of 7604 Nastaliq Urdu Facebook comments, extending an earlier corpus by correcting text-level inconsistencies and re-examining the original labels. Second, we fine-tune four Urdu-capable transformers under a common grid-search protocol and compare them against twelve feature-classifier baselines on identical data partitions. The best configuration, Multilingual Representations for Indian Languages (MuRIL), reaches 87.05% accuracy and 87.06% macro F1, improving on the strongest baseline by 5.52 percentage points, a margin that exceeds the 95% confidence interval on test accuracy; the two leading transformers differ by less than the test set can resolve, and performance is stable across five optimizers. Third, we compute word-level SHAP attributions adapted to right-to-left rendering. Attribution mass concentrates on high-frequency slurs but extends to syntactically neutral function words and to political vocabulary characteristic of the source pages, indicating that aggregate performance may rest partly on lexical and topical regularities rather than a generalizable representation of offensiveness.