Enhancing Arabic aspect category detection using large language models (LLMs)

Large language models (LLMs) have revolutionized the world of natural language processing (NLP) by breaking through a variety of tasks, such as text generation, translation, and text summarization. In this study, we tackle aspect-based sentiment analysis (ABSA), a subtask of sentiment analysis that concentrates on gathering and collecting aspect terms and their corresponding sentiment polarity. ABSA can be divided into three subtasks: aspect term extraction, aspect sentiment classification, and aspect category detection (ACD). While most research has focused on the English language, exploring other low-resource languages such as Arabic remains limited and challenging. Therefore, the purpose of this paper is to provide insight into Arabic aspect category detection using zero-shot prompts, few-shot prompts, chain-of-thought prompts, combination of prompting and fine-tuning of large language models (LLMs). Moreover, we apply this approach to two datasets: the Gaza Reviews and Book Reviews. The experiment illustrates that our proposed method significantly outperforms the baseline and related work models, achieving an enhanced F1-score of over 10 percentage points for the Arabic news dataset and an improved F1-score of over 17 percentage points for the Book Reviews (HAAD) dataset.