Novel approach for Arabic fake news classification using embedding from large language features with CNN-LSTM...

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of fake news necessitates robust information quality management.
• Existing fake news detection methods face limitations, particularly for Arabic text.

Purpose of the Study:

• To develop and evaluate an advanced framework for fake news detection in Arabic.
• To overcome limitations of current approaches using deep learning and ensemble methods.

Main Methods:

• Employed deep learning models including CNN, LSTM, EfficientNetB4, Inception, Xception, ResNet, and ConvLSTM.
• Proposed a novel voting ensemble framework combining CNN and LSTM.
• Integrated ELMo word embeddings, comparing its contextual representation with GloVe, BERT, FastText, and FastText subwords.
• Utilized LIME XAI technique for feature contribution insights and 5-fold cross-validation.

Main Results:

• The proposed voting ensemble with ELMo embeddings achieved superior performance, with accuracy (98.42%), precision (98.54%), recall (99.5%), and F1-score (98.93%).
• The framework demonstrated competitive performance against state-of-the-art transformer architectures like BERT and RoBERTa.
• Achieved reduced inference time and enhanced interpretability compared to transformer models.

Conclusions:

• The developed framework offers an efficient and highly effective solution for Arabic fake news detection.
• The research provides a reliable and interpretable text classification solution, advancing the field of fake news detection.
• Findings highlight the potential for improved information management and combating misinformation.