Sentiment analysis for deepfake X posts using novel transfer learning based word embedding and hybrid LGR approach

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Social Media Analysis

Background:

• Social media fuels content sharing, including AI-generated deepfake text used for misinformation.
• Existing sentiment analysis struggles with the nuances of deepfake content, failing to capture its misleading nature.
• Accurately understanding public opinion on deepfake posts is increasingly challenging.

Purpose of the Study:

• To propose a hybrid deep learning (DL) and transfer learning (TL) approach for sentiment analysis of deepfake posts.
• To develop novel TL-based features combining LSTM and DT for enhanced contextual understanding.
• To evaluate the proposed method against various machine learning (ML) and DL techniques.

Main Methods:

• A hybrid DL model incorporating LSTM, GRU, and RNN (LGR) was developed.
• Novel transfer learning (TL) features were engineered by combining LSTM and Decision Tree (DT).
• Models were trained using diverse feature extraction techniques (BOW, TF-IDF, word embeddings) and validated with k-fold cross-validation.

Main Results:

• The proposed LGR approach with novel TL features achieved 99% accuracy in sentiment analysis.
• The hybrid DL and TL method outperformed existing ML and DL techniques.
• Extensive hyperparameter tuning enhanced ML model performance and efficiency.

Conclusions:

• The developed approach effectively detects and mitigates the spread of deepfake content.
• This study provides a scalable mechanism for monitoring and reducing the impact of online misinformation.
• The findings address a critical gap in analyzing deepfake-specific social media sentiment.