A Staged Framework for Mining User Requirements from Vehicle Online Reviews Based on Large Language Models and Sentiment Weighting

Large-scale online reviews provide a valuable source of user feedback, yet existing methods still offer limited support for transforming massive review corpora into structured and decision-relevant requirement knowledge. Rule-based and manually intensive approaches are difficult to scale, while direct end-to-end use of large language models often faces challenges in maintaining stable requirement structures and practical large-scale deployment. To address these limitations, this study proposes a staged framework for mining user requirements from vehicle online reviews, with a particular focus on supporting early-stage requirement engineering. The framework uses large language models for requirement taxonomy construction and automatic annotation and transfers large-scale requirement categorization to a BERT classifier to balance semantic capability with deployment efficiency. A mini-batch iterative strategy is further introduced to progressively induce requirement categories from review data rather than fully predefining them in advance. In addition, sentiment weighting is incorporated to prioritize requirement categories and better reflect user pain points in subsequent analysis and decision support. Experiments on 467,962 review texts show that the proposed method outperforms several conventional machine learning and deep learning baselines on the studied dataset. Beyond quantitative evaluation, the study also examines the structural characteristics of online review-based requirement identification and explores the applicability of the framework in other review scenarios. Overall, the proposed framework provides a practical systems-oriented workflow for large-scale requirement analysis and contributes a review-based approach to early-stage requirement engineering, including requirement identification, categorization, prioritization, and interpretation.