Abstract
The analysis of Arabic Twitter data sets is a highly active research topic, particularly since the outbreak of COVID-19 and subsequent attempts to understand public sentiment related to the pandemic. This activity is partially driven by the high number of Arabic Twitter users, around 164 million. Word embedding models are a vital tool for analysing Twitter data sets, as they are considered one of the essential methods of transforming words into numbers that can be processed using machine learning (ML) algorithms. In this work, we introduce a new model, Arab2Vec, that can be used in Twitter-based natural language processing (NLP) applications. Arab2Vec was constructed using a vast data set of approximately 186,000,000 tweets from 2008 to 2021 from all Arabic Twitter sources. This makes Arab2Vec the most up-to-date word embedding model researchers can use for Twitter-based applications. The model is compared with existing models from the literature. The reported results demonstrate superior performance regarding the number of recognised words and F1 score for classification tasks with known data sets and the ability to work with emojis. We also incorporate skip-grams with negative sampling, an approach that other Arabic models haven't previously used. Nine versions of Arab2Vec are produced; these models differ regarding available features, the number of words trained on, speed, etc. This paper provides Arab2Vec as an open-source project for users to employ in research. It describes the data collection methods, the data pre-processing and cleaning step, the effort to build these nine models, and experiments to validate them qualitatively and quantitatively.