A massive amount of textual data is shared each day through the different social media platforms. Analyzing this data to understand the sentiment of users (positive, negative, natural) about various topics and products is a challenging task when it comes to analyze Arabic text due to the morphological diversity and complex grammatical structure of this language, and is more challenging when analyzing social media text with the great diversity in writing style and dialects. Due to the lack of Arabic resources for Sentiment Analysis and the increasing need to create dialectical resources. In this paper, we present a Syrian dialectical Arabic dataset consists of 20k Facebook comment that are manually annotated for three classes positive, negative and natural. We used finetuned AraBERT as our baseline to evaluate our corpus. Also, compared its performance with many traditional machine learning classifiers, LSTM and MARBERT models. The experimental results show that the finetuned AraBERT outperformed all the other algorithms with f1-score of 81%.
Ahmad et al. (Mon,) studied this question.