Arab2Vec: An Arabic word embedding model for use in Twitter NLP applications

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning (ML)
• Computational Linguistics

Background:

• Arabic Twitter data analysis is crucial for understanding public sentiment, especially post-COVID-19.
• Word embedding models are essential for transforming text data into numerical formats for ML algorithms.
• Existing Arabic word embedding models have limitations in scope and functionality.

Purpose of the Study:

• To introduce Arab2Vec, a novel and up-to-date word embedding model specifically designed for Arabic Twitter data.
• To enhance natural language processing applications on Arabic social media.
• To provide a superior alternative to existing Arabic word embedding models.

Main Methods:

• Construction of Arab2Vec using a large dataset of approximately 186 million Arabic tweets (2008-2021).
• Implementation of skip-grams with negative sampling, a novel approach for Arabic models.
• Development of nine distinct Arab2Vec model versions with varying features and training parameters.

Main Results:

• Arab2Vec demonstrates superior performance compared to existing models in terms of recognized words and F1 score for classification tasks.
• The model exhibits effective handling of emojis within Arabic text.
• Validation through qualitative and quantitative experiments confirms the model's efficacy.

Conclusions:

• Arab2Vec represents a significant advancement in Arabic word embedding models for Twitter data.
• The open-source release of Arab2Vec facilitates further research and application in NLP.
• The model's enhanced capabilities offer improved insights into Arabic social media discourse.