Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fake News Classification in Urdu: A Domain Adaptation Approach for a Low-Resource Language

Muhammad Zain Ali, Bernhard Pfahringer, Tony Smith

TL;DR

This work tackles misinformation detection in Urdu, a low-resource language, by applying domain-adaptive pretraining to two multilingual transformers (XLM-RoBERTa and mBERT) using a large Urdu news corpus before downstream fine-tuning on four Urdu fake-news datasets. The authors demonstrate that domain-adapted XLM-R consistently improves classification performance across datasets, while domain-adapted mBERT yields mixed results, despite substantial perplexity reductions. The findings highlight the importance of model choice in domain adaptation for low-resource languages and provide a replicable framework for enhancing Urdu NLP tasks through domain-aware pretraining and careful fine-tuning. Overall, the study advances fake-news detection in Urdu and offers practical guidance for deploying domain-adaptive transformer models in low-resource settings.

Abstract

Misinformation on social media is a widely acknowledged issue, and researchers worldwide are actively engaged in its detection. However, low-resource languages such as Urdu have received limited attention in this domain. An obvious approach is to utilize a multilingual pretrained language model and fine-tune it for a downstream classification task, such as misinformation detection. However, these models struggle with domain-specific terms, leading to suboptimal performance. To address this, we investigate the effectiveness of domain adaptation before fine-tuning for fake news classification in Urdu, employing a staged training approach to optimize model generalization. We evaluate two widely used multilingual models, XLM-RoBERTa and mBERT, and apply domain-adaptive pretraining using a publicly available Urdu news corpus. Experiments on four publicly available Urdu fake news datasets show that domain-adapted XLM-R consistently outperforms its vanilla counterpart, while domain-adapted mBERT exhibits mixed results.

Fake News Classification in Urdu: A Domain Adaptation Approach for a Low-Resource Language

TL;DR

This work tackles misinformation detection in Urdu, a low-resource language, by applying domain-adaptive pretraining to two multilingual transformers (XLM-RoBERTa and mBERT) using a large Urdu news corpus before downstream fine-tuning on four Urdu fake-news datasets. The authors demonstrate that domain-adapted XLM-R consistently improves classification performance across datasets, while domain-adapted mBERT yields mixed results, despite substantial perplexity reductions. The findings highlight the importance of model choice in domain adaptation for low-resource languages and provide a replicable framework for enhancing Urdu NLP tasks through domain-aware pretraining and careful fine-tuning. Overall, the study advances fake-news detection in Urdu and offers practical guidance for deploying domain-adaptive transformer models in low-resource settings.

Abstract

Misinformation on social media is a widely acknowledged issue, and researchers worldwide are actively engaged in its detection. However, low-resource languages such as Urdu have received limited attention in this domain. An obvious approach is to utilize a multilingual pretrained language model and fine-tune it for a downstream classification task, such as misinformation detection. However, these models struggle with domain-specific terms, leading to suboptimal performance. To address this, we investigate the effectiveness of domain adaptation before fine-tuning for fake news classification in Urdu, employing a staged training approach to optimize model generalization. We evaluate two widely used multilingual models, XLM-RoBERTa and mBERT, and apply domain-adaptive pretraining using a publicly available Urdu news corpus. Experiments on four publicly available Urdu fake news datasets show that domain-adapted XLM-R consistently outperforms its vanilla counterpart, while domain-adapted mBERT exhibits mixed results.
Paper Structure (28 sections, 1 equation, 11 figures, 3 tables)

This paper contains 28 sections, 1 equation, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Fake News Detection
  4. Domain Adaptation
  5. Methodology
  6. Datasets
  7. Urdu News Articles Dataset
  8. Fake News Datasets
  9. Proposed Architecture
  10. Domain Adaptation Process
  11. Fine-tuning the Model
  12. Experimental Setup
  13. Model Training
  14. Masked Language Model
  15. XLM-R model
  16. ...and 13 more sections

Figures (11)

  • Figure 1: Proposed Architecture. (left side) Basic fine-tuning; (right side) Domain Adaptation + Fine-Tuning. PLM = Pre-trained Language Model.
  • Figure 2: Domain Adaptation (Masked Language Model) Loss Curves for XLM-R model
  • Figure 3: Domain Adaptation (Masked Language Model) Loss Curves for mBERT model
  • Figure 4: Loss and Accuracy curves for Classifier trained on Ax-to-Grindharris2023ax using XLM-R models
  • Figure 5: Loss and Accuracy curves for Classifier trained on UFN2023farooq2023fake using XLM-R models
  • ...and 6 more figures