AraCust: a Saudi Telecom Tweets corpus for sentiment analysis.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Sentiment Analysis

Background:

• Arabic language resources for NLP are scarce compared to other languages.
• Existing Arabic corpora often rely on translation, limiting their authenticity.
• Dialectal Arabic presents unique challenges for NLP tasks like sentiment analysis.

Purpose of the Study:

• To construct and present AraCust, a novel 200,000-tweet Gold Standard Corpus for Dialectal Arabic Sentiment Analysis.
• To provide a high-quality, manually annotated dataset specifically for Saudi dialect tweets.
• To facilitate research and development in Arabic Sentiment Analysis.

Main Methods:

• Collection and curation of a large Arabic tweets dataset.
• Rigorous cleaning, pre-processing, and manual annotation for sentiment analysis (inter-annotator agreement k=0.60).
• Exploratory data analysis to understand corpus features and guide ASA method selection.

Main Results:

• The AraCust corpus demonstrates superior performance compared to existing datasets.
• A supervised classifier trained on AraCust achieved 91% accuracy and 89% F1avg score.
• Exploratory analysis identified key features relevant to Saudi dialect sentiment.

Conclusions:

• AraCust is a valuable, high-quality resource for advancing Arabic Sentiment Analysis, particularly for dialectal variations.
• The corpus enables benchmark evaluations and the development of more effective ASA models.
• AraCust and associated code will be publicly released on GitHub to support the research community.