ArabBert-LSTM: improving Arabic sentiment analysis based on transformer model and Long Short-Term Memory.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Sentiment analysis automates opinion mining from textual data across platforms like social media.
• Arabic text presents unique morphological complexities that challenge sentiment analysis.
• Existing methods often struggle to capture the nuances of Arabic language for accurate sentiment classification.

Purpose of the Study:

• To propose and evaluate a deep learning model for enhanced Arabic sentiment analysis.
• To address the morphological intricacies of the Arabic language in sentiment classification.
• To improve the accuracy and reliability of sentiment analysis for Arabic text.

Main Methods:

• A hybrid deep learning model combining Arabert (Transformer-based Model for Arabic Language Understanding) for word embeddings and Long Short-Term Memory (LSTM) for sequence modeling.
• Utilizing feedforward neural networks and an output layer for classification.
• Comparison with traditional machine learning and other deep learning algorithms using various vectorization techniques (TF-IDF, ArabBert, CBOW, skipGrams) on four Arabic datasets.

Main Results:

• The proposed Arabert-LSTM model significantly improved sentiment analysis accuracy compared to baseline methods.
• Achieved an accuracy rate exceeding 97% on Arabic sentiment analysis tasks.
• Demonstrated the effectiveness of transformer models and LSTM in capturing contextual information and long-term dependencies in Arabic text.

Conclusions:

• The developed deep learning framework offers a robust solution for Arabic sentiment analysis.
• Leveraging transformer models like Arabert and sequence modeling with LSTM is highly effective for Arabic text.
• The research advances the field of Arabic sentiment analysis, providing a more accurate and reliable tool for opinion mining.