Predicting customer sentiment: the fusion of deep learning and a fuzzy system for sentiment analysis of Arabic text

• Published in: SOCIAL NETWORK ANALYSIS AND MINING
• Main Authors: Ambreen, Shela; Iqbal, Muhammad; Asghar, Muhammad Zubair; Mazhar, Tehseen; Khattak, Umar Farooq; Khan, Muhammad Amir; Hamam, Habib
• Format: Article
• Language: English
• Published: SPRINGER WIEN 2024
• Subjects: Computer Science
• Online Access: https://www-webofscience-com.uitm.idm.oclc.org/wos/woscc/full-record/WOS:001335841500001

Understanding client feedback and satisfaction is a critical concern for any business organization operating in the highly competitive internet industry. Notably, social media platforms such as X (Twitter) act as forums for customers to voice their opinions. Analyzing such feedback is beneficial since it provides insights into client interests. The proposed model addresses various challenges, such as measuring customer satisfaction levels from Arabic text by proposing a hybrid deep learning technique enriched with fuzzy logic. The proposed system aims to construct an Arabic sentiment-based system that uses an innovative combination of fuzzy logic and a deep neural network to evaluate customer satisfaction, hence assisting businesses in improving their service and product quality. To forecast sentiment polarity (positive or negative), the proposed method employs bidirectional long short-term memory (LSTM) with an attention component. Following that, the level of consumer contentment is determined using fuzzy logic. Ablation studies demonstrate the importance of the attention mechanism, which contributes to a considerable improvement in accuracy compared to a BiLSTM-only model. Fuzzy logic incorporation increases the ability of a model to handle imprecision and uncertainty in sentiment formulations, helping it to additionally correct sentiment analysis. Furthermore, hyperparameter adjustment improves performance by highlighting the model's sensitivity to specific variables. The system achieved an excellent accuracy of 95%, outperforming earlier baseline techniques. Furthermore, the efficacy of the suggested approach was demonstrated using statistical testing.