Empirical Analysis for Arabic Target-Dependent Sentiment Classification using LLMs

This work aims to develop a valuable tool for understanding sentiments on specific topics in the Arabic language. Currently, most sentiment analysis research focuses on social media platforms like X (formerly known as Twitter), which provide a rich source of information and opinions but lack the precise tools required for accurate Arabic sentiment analysis. This limitation has hindered the development of a reliable and robust sentiment analysis model for Arabic. To address this challenge, we present a technique that emphasizes target-based sentiment classification using an advanced approach based on a Large Language Model (LLM). The presented model is trained on the AT-ODTSA dataset. This dataset includes manually classified Arabic tweets along with identified topics (targets) and sentiments. By leveraging this dataset, our work shows enhancement with the sentiment classification for Arabic tweets. The presented technique achieved a classification accuracy of 0.76 for target-independent sentiment classification and 0.77 to classify sentiment dependent on a target. The method that is being offered uses a fine-tuned version of the Arabic-MARBERT-sentiment model. Our work is expected to contribute to a deeper and more accurate understanding of public opinions on specific topics in the Arab world.