KD-BERT: A Knowledge-Driven BERT Framework for Few-Shot Requirement Classification and Kano-Based Prioritization

Rapid release cycles and user-centric design have transformed requirement engineering into a continuous process within software development. Analyzing user-generated feedback from diverse online platforms is essential for feature prioritization. While frameworks like the Kano model offer a structured approach to understanding user satisfaction, traditional methods based on surveys and interviews are inefficient and not scalable for large datasets. Consequently, there is growing interest in automating this analysis using natural language processing (NLP) models. However, these models typically require large amounts of labeled data, creating the “few-shot problem,” particularly in dynamic domains where user needs change and annotated data are limited. To address these challenges, we developed KD-BERT, a knowledge-driven framework that augments a multi-input transformer-based architecture with generative artificial intelligence and ontology-based reasoning. Our model uses semantic metrics to estimate user satisfaction and achievement scores, providing a data-driven method for Kano category prediction and prioritization. The KD-BERT framework was evaluated using mobile banking user reviews, representing a FinTech domain. Its ontology-driven design provides a foundation for adaptation through domain-specific knowledge. KD-BERT showed progressive numerical gains over the baseline: semantic metrics increased Macro-F1 from 0.835 to 0.844, although this difference was not statistically significant, while further enhancements achieved Macro-F1 of 0.954.