Transfer learning driven fake news detection and classification using large language models.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Social Science

Background:

• The rapid spread of fake news on social media poses significant societal and individual challenges.
• Current fake news detection methods struggle with limited data, linguistic complexity, and embedding integration.

Purpose of the Study:

• To develop a robust fake news detection framework for limited data scenarios.
• To enhance the accuracy and interpretability of fake news classification models.

Main Methods:

• Proposed a multi-stage transfer learning framework using RoBERTa, a pre-trained large language model.
• Systematically compared word embedding techniques (Word2Vec, one-hot encoding) with refined fine-tuning.
• Evaluated the framework on two real-world benchmark datasets.

Main Results:

• Achieved a minimum accuracy improvement of 3.9% over state-of-the-art models.
• Demonstrated the effectiveness of the transfer learning approach in limited data settings.
• Provided insights into the impact of embedding techniques on fake news classification.

Conclusions:

• The novel multi-stage transfer learning framework offers a more robust and accurate solution for fake news detection.
• The approach effectively addresses limitations of existing methods, particularly in low-data environments.
• This work highlights the importance of embedding techniques and transfer learning for reliable fake news identification.