Improving Arabic Sentiment Analysis Using CNN-Based Architectures and Text Preprocessing.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Sentiment analysis is crucial for numerous natural language applications.
• Arabic sentiment analysis presents unique challenges due to language complexities.

Purpose of the Study:

• To evaluate two deep learning models (CNN and CNN-BiGRU) for Arabic sentiment analysis.
• To improve upon existing benchmarks for Arabic sentiment classification on the ASTD and ATDFS datasets.
• To investigate the impact of novel Arabic preprocessing techniques.

Main Methods:

• Application of a 2-layer CNN (MC1) and a 2-layer CNN with BiGRU (MC2) to Arabic text datasets.
• Experiments conducted on both 2-class and multiclass sentiment analysis tasks.
• Detailed analysis of Arabic preprocessing, including emoticon processing and custom stoplists.

Main Results:

• Achieved 73.17% on the difficult ASTD 4-class task, surpassing previous work (65.58%).
• Reached 90.06% on the 2-class task using MC1, outperforming prior results (85.58%).
• Preprocessing steps like emoticon handling and custom stoplists yielded significant performance gains (up to 5.48%).

Conclusions:

• The proposed models and preprocessing techniques demonstrate superior performance for Arabic sentiment analysis.
• Novel preprocessing steps are vital for enhancing accuracy in Arabic NLP tasks.
• The study provides valuable insights for developing more effective Arabic sentiment analysis systems.