Addressing Data Scarcity in Bangla Fake News Detection: An LLM-Based Dataset Augmentation Approach

2026-05-05 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, extended

Summary

This study introduces an LLM-based data augmentation framework to enhance fake news detection in Bangla, a low-resource language, addressing limitations of small and imbalanced datasets like BanFakeNews. Researchers used the instruction-tuned Gemma-3-27B-IT model to generate 4,545 synthetic Bangla fake news articles. The framework employs semantic filtering and controlled subsampling to maintain label consistency and diversity. Experiments compared zero-shot and few-shot prompting, multiple augmentation rates, and random versus similarity-based selection strategies. The most effective configuration involved augmenting only the minority fake news class with a high augmentation rate (K=5) and random subsampling using zero-shot prompting, which improved the Fake News F1 score from 0.8560 to 0.8800, a 2.4-point gain. The generated dataset and implementation are publicly released to support further research.

Key takeaway

For AI Engineers and Research Scientists working on natural language processing in under-resourced languages, this research demonstrates that carefully controlled LLM-based data augmentation can significantly boost model performance. You should prioritize augmenting only the minority class, using zero-shot prompting, and employing random subsampling at higher augmentation rates to maximize linguistic diversity and improve fake news detection F1 scores. This approach offers a practical path to overcome data scarcity challenges.

Key insights

LLM-based data augmentation significantly improves fake news detection in low-resource languages by generating diverse synthetic data.

Principles

• Targeted minority class oversampling is effective.
• Linguistic diversity in synthetic data enhances generalization.
• Zero-shot prompting can outperform few-shot in low-resource settings.

Method

The proposed framework generates synthetic news using Gemma-3-27B-IT with zero-shot prompting, followed by random subsampling of K=5 articles to augment only the minority fake news class.

In practice

• Augment only minority classes for imbalanced datasets.
• Prioritize random subsampling over similarity-based for diversity.
• Consider zero-shot prompting for LLM-based augmentation in low-resource NLP.

Topics

• Bangla Fake News Detection
• LLM Data Augmentation
• Gemma-3-27B-IT
• Zero-Shot Prompting
• Class Imbalance

Code references

• phigratio/bangla-fake-news

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Label-Aware Pseudo-Training Sample Generation for Text Classification — A new data augmentation method uses LLMs to subtly alter sentences, improving NLP performance with limited data.
• Retrieval-Augmented Generation with Small Language Models for Fake News Detection — RAG-based fake news detection using web searches did not outperform traditional or deep learning methods.
• LLM-based Detection of Manipulative Political Narratives — A framework combines prompt-based filtering with unsupervised clustering to detect and structure manipulative political narratives.
• Fact4ac at the Financial Misinformation Detection Challenge Task: Reference-Free Financial Misinformation Detection via Fine-Tuning and Few-Shot Prompting of Large Language Models — Detecting financial misinformation without external references is achievable using LLMs combined with specific fine-tuning.
• In-Context Learning for the Imputation of Public Opinion Data with Large Language Models — In-context learning with LLMs significantly improves missing public opinion data imputation over traditional statistical methods.
• Scikit-LLM vs. Traditional Text Classifiers: When Should You Use an LLM? — LLMs offer superior zero-shot text classification accuracy and competitive latency, especially with limited data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.