Sentiment Analysis of Customers in Iran’s Automotive Industry Using the LSTM Method

Abstract

Given the large volume of customer opinions and complaints regarding automobiles in Iranian virtual platforms, this study aimed to develop an accurate sentiment analysis model based on Long Short-Term Memory (LSTM) neural networks for processing Persian texts. The primary objective was to improve the accuracy of sentiment classification (positive, negative, and neutral) by overcoming the challenges associated with imbalanced data and linguistic limitations. Data were collected from 12 valid online sources, including 4,348 customer comments regarding various automobile models. Three LSTM architectures (single-layer, two-layer, and three-layer) were implemented and compared using fixed parameters (Embedding = 128, Hidden Units = 64, Dropout = 0.3). To enhance performance, Data Augmentation, the SMOTE algorithm, and FastText word embeddings were employed. The three-layer LSTM model outperformed the shallower models, achieving a validation accuracy of 62.01%, which increased to 95.45% after applying optimization techniques. The two-layer model demonstrated superior performance in identifying positive comments, whereas the three-layer model showed a more balanced performance across all sentiment classes. The issue of overfitting in simpler models was mitigated through the addition of layers and the use of Dropout. The findings indicated that the two-layer model improved overall accuracy, particularly in recognizing positive sentiments, compared to the single-layer model. Furthermore, the three-layer LSTM model demonstrated greater balance and robustness across all classes. To further improve performance, the three-layer model was retrained using Data Augmentation techniques and the SMOTE algorithm for data balancing. The modified version substantially improved overall accuracy and validation performance.