Analyzing customer perceptions has become increasingly important in the automotive industry as it provides actionable insights into consumer satisfaction, preferences, and areas requiring improvement. This study proposes a novel perception analysis framework for automotive reviews using advanced Large Language Models (LLMs), including BERT, FLAN-XXL, and Mistral 7B, leveraging zero-shot learning to categorize reviews without task-specific training data. The framework follows a two-stage evaluation process, beginning with zero-shot perception classification and followed by a detailed topic-wise perception analysis. Model performance was evaluated using accuracy, precision, recall, and F1-score across reviews from five major automotive brands—Toyota, Kia, Honda, Nissan, and Hyundai. While BERT and FLAN-XXL achieved high accuracy in certain cases, they exhibited limitations in balancing precision and recall, particularly for nuanced perceptions. In contrast, Mistral 7B consistently demonstrated superior F1-score performance and was therefore selected for in-depth topic-wise analysis across key automotive aspects, including comfort, safety, performance, design, and price. The results indicate that Mistral 7B excels in objective categories such as performance and price, while performance declines for more subjective topics such as comfort and safety. Overall, the findings highlight the effectiveness of zero-shot LLM-based perception analysis for automating large-scale customer feedback analysis in the automotive domain. Future work will focus on domain-specific fine-tuning and the integration of multimodal information to further improve performance, particularly in subjective perception categories, enabling more accurate and data-driven decision-making in the automotive industry. • Leverage large language models for automotive customer perception analysis. • Perform zero-shot and topic-wise perception analysis on car reviews. • Demonstrate the effectiveness of large language models in nuanced perception analysis.
Mathew et al. (Thu,) studied this question.