Eliciting Explicit and Implicit Requirements from App Reviews via Grounded Theory and BERT-Based Classification

Context: App reviews contain valuable user requirements—some stated explicitly, others implied within narratives of use—yet prior work has largely focused on explicit feedback alone. Objective: We investigate the characterization and automatic identification of both explicit and implicit requirements-relevant signals from app reviews. Method: We applied grounded theory to derive a classification scheme with 14 requirements-relevant characteristics distinguishing explicit, implicit, and irrelevant reviews. Five RE/SE-experienced annotators labeled 84,721 reviews across 20 apps and 13 categories, producing the REV4RE gold-standard dataset (κ = 0.74). We then fine-tuned four BERT-based models across imbalanced, class-weighted, and under-sampled configurations. Results: RoBERTa achieved a weighted F1-score of 87%, numerically comparable to human annotator performance. Random under-sampling improved implicit requirements detection by approximately 40%, with AUC scores of 0.91–0.92. Conclusions: Our findings demonstrate the feasibility of automated extraction of both explicit and implicit requirements from app reviews at scale, advancing review-based requirements elicitation by recognizing both as distinct categories of requirements-relevant signals.