REAM: A Multi-Sentiment Review-Driven Method for Fine-Grained Customer Requirement Extraction and Analysis

Massive online reviews play a crucial role in mining customer requirements. However, existing research has three limitations: (1) the inadequate capture of contextual semantics in traditional text classification models, (2) the lack of sentiment analysis targeted at the product attribute level, and (3) the neglect of inter-attribute dependencies in requirement analysis. To address these gaps, this study proposed a requirement extraction and analysis method (REAM) that integrated multi-sentiment reviews. Specifically, a multi-domain adaptive BERT–BiLSTM–hierarchical position-aware network (HPANet) was developed for the fine-grained classification of reviews to automatically categorize them into nine key product attributes (for example, cost-performance ratio and appearance) that served as concrete manifestations of customer requirements. For requirements analysis, this method examined three dimensions: customer attention to product features, sentiment orientation toward specific attributes, and inter-attribute dependencies, thereby facilitating a comprehensive understanding of customer requirements. In the requirement extraction phase, the experimental results demonstrated that the proposed model achieved 1.2% higher classification accuracy than the conventional BERT model on a custom-built customer requirement dataset. To further verify the stability of the BERT–BiLSTM–HPANet model, additional validation was conducted on public datasets (CNews and ChnSentiCorp), revealing significant improvements in both the accuracy and F1-score compared with the baseline models.