Ensemble learning approach for distinguishing human and computer-generated Arabic reviews.

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Customer reviews are vital for businesses, but AI-generated fake reviews erode consumer trust.
• Existing research on detecting AI-generated text primarily focuses on English, with a gap in Arabic.
• Ensemble learning (EL) techniques for classifying Arabic fake reviews are underexplored.

Purpose of the Study:

• To develop and evaluate models for classifying human versus computer-generated Arabic reviews.
• To investigate the effectiveness of ensemble learning, specifically soft voting, in this classification task.
• To identify linguistic features that differentiate human and AI-generated Arabic reviews.

Main Methods:

• Employed traditional machine learning, deep learning, and transformer models.
• Utilized ensemble techniques, including soft voting, combining logistic regression (LR) and convolutional neural network (CNN) models.
• Conducted textual analysis focusing on parts of speech (POS), emotions, and linguistic patterns.

Main Results:

• Achieved high accuracy in classification, with an ensemble of LR and CNN reaching 89.70%, comparable to AraBERT's 90.0%.
• Identified significant linguistic disparities: AI reviews contained a much higher proportion of adjectives (6.3%) than human reviews (0.46%).
• Demonstrated the effectiveness of ensemble methods in improving fake review detection.

Conclusions:

• The study successfully differentiates human and AI-generated Arabic reviews, offering a valuable tool for businesses.
• Linguistic analysis provides key insights into the characteristics of fake reviews, aiding detection efforts.
• Advances Arabic Natural Language Processing (NLP) and provides practical implications for maintaining marketplace integrity and consumer trust.