Evaluating Arabic Emotion Recognition Task Using ChatGPT Models: A Comparative Analysis between Emotional Stimuli Prompt, Fine-Tuning, and In-Context Learning

Textual emotion recognition (TER) has significant commercial potential since it can be used as an excellent tool to monitor a brand/business reputation, understand customer satisfaction, and personalize recommendations. It is considered a natural language processing task that can be used to understand and classify emotions such as anger, happiness, and surprise being conveyed in a piece of text (product reviews, tweets, and comments). Despite the advanced development of deep learning and particularly transformer architectures, Arabic-focused models for emotion classification have not achieved satisfactory accuracy. This is mainly due to the morphological richness, agglutination, dialectal variation, and low-resource datasets of the Arabic language, as well as the unique features of user-generated text such as noisiness, shortness, and informal language. This study aims to illustrate the effectiveness of large language models on Arabic multi-label emotion classification. We evaluated GPT-3.5 Turbo and GPT-4 using three different settings: in-context learning, emotional stimuli prompt, and fine-tuning. The ultimate objective of this research paper is to determine if these LLMs, which have multilingual capabilities, could contribute to enhancing the aforementioned task and encourage its use within the context of an e-commerce environment for example. The experimental results indicated that the fine-tuned GPT-3.5 Turbo model achieved an accuracy of 62.03%, a micro-averaged F1-score of 73%, and a macro-averaged F1-score of 62%, establishing a new state-of-the-art benchmark for the task of Arabic multi-label emotion recognition.