Abstract
Misinformation in low-resource languages like Bangla presents a significant challenge due to linguistic complexity and a lack of specialized tools. This study proposes a deep learning framework utilizing Bidirectional Gated Recurrent Units (Bi-GRU) to detect fake news. Using a dataset of approximately 50,000 articles, the proposed model achieved a state-of-the-art accuracy of 99.16%, significantly outperforming traditional machine learning and standard sequential models.
Introduction
While fake news detection is a mature field for high-resource languages like English, Bangla—spoken by over 200 million people—remains underserved in the AI community. The language's rich idiomatic expressions and diverse dialects make traditional keyword-based detection ineffective. This research addresses the gap by developing a robust pipeline specifically tuned for the unique syntax and context of Bangladeshi news media.
Methodology
The research implemented a comprehensive pipeline to process and classify news articles. The workflow involves data preprocessing, visualization, text cleaning, and model training.
Model Architecture
The core of the proposed system is a Bidirectional Gated Recurrent Unit (Bi-GRU). Unlike standard GRUs, the bidirectional variant processes text in both forward and backward directions, capturing context from both the preceding and succeeding words. This is particularly effective for Bangla, where sentence veracity often depends on holistic context.
The architecture includes an embedding layer to convert tokens into dense vectors, followed by the Bi-GRU layer to capture temporal dependencies. Global Average Pooling is used to reduce dimensionality, while Dense and Dropout layers provide the final classification and regularization to prevent overfitting.
Data and Preprocessing
The study utilized a dataset sourced from Kaggle consisting of approximately 50,000 news items, including Authentic-48K.csv and Fake-1K.csv. To address the significant class imbalance, the researchers employed resampling techniques. Preprocessing was a critical step, involving the removal of line breaks, tabs, and specific Bengali punctuation, alongside using a specialized "Stopwords Bengali (BN)" library to filter linguistic noise.
Results and Discussion
The Bi-GRU model demonstrated exceptional performance, setting a new benchmark for Bangla NLP. The results were remarkably high across all primary metrics:
- Accuracy: 99.16%
- Precision: 99.40%
- Recall: 99.74%
- F1-Score: 99.57%
Comparatively, standard LSTM models achieved 97.49% accuracy, while 1D CNNs reached 99.06%. The results suggest that Bi-GRU's computational efficiency and bidirectional context-handling are ideal for this specific linguistic task, surpassing previous efforts that typically reached the 91-96% range.
Conclusion
This study confirms that deep learning, specifically Bi-GRU architectures, can effectively combat misinformation in low-resource languages. Future work will focus on expanding the dataset with "pseudo-events" to ensure robustness against evolving misinformation tactics and developing real-time automation tools, such as browser extensions, to protect citizens from digital misinformation.
Summary updated