Hybrid transformer-fuzzy framework for interpretable sentiment classification in deepfake social media content.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Social Science

Background:

• Deepfake text on social media poses a significant challenge to discerning genuine user sentiment.
• Existing sentiment analysis models often lack interpretability and generalizability due to reliance on black-box feature engineering.
• There is a critical need for robust and explainable frameworks for analyzing online text.

Purpose of the Study:

• To develop and evaluate a novel hybrid framework for detecting AI-generated (deepfake) text on social media.
• To enhance the interpretability and generalizability of sentiment analysis models for manipulative text.
• To address the challenge of distinguishing genuine user attitudes from artificial narratives in high-volume online data.

Main Methods:

• A hybrid approach combining Transformer-based contextual embeddings for semantic richness.
• Integration of a Fuzzy Rule-Based System (FRBS) for human-readable explanations.
• Utilized the balanced TweepFake dataset (25,572 tweets) for evaluation under 10-fold cross-validation.

Main Results:

• The proposed hybrid framework achieved 97.0% accuracy and F1-score.
• Demonstrated stable performance across epochs and superior explainability compared to baseline models.
• Outperformed the RoBERTa baseline by 1.6% in accuracy.

Conclusions:

• The hybrid model offers a robust and explainable solution for deepfake text detection in sentiment analysis.
• Integrating contextual modeling with symbolic reasoning enhances model interpretability and trustworthiness.
• This framework advances the ability to analyze user attitudes towards manipulative online content.