LePB-SA4RE: A Lexicon-Enhanced and Prompt-Tuning BERT Model for Evolving Requirements Elicitation from App Reviews

Pre-trained language models with fine-tuning (FT) have achieved notable success in aspect-based sentiment analysis (ABSA) for automatic requirements elicitation from app reviews. However, the fixed parameters during FT progress often face challenges when applied to low-resource and noisy app review scenarios. Although prompt-tuning (PT) has gained attention in ABSA for its flexibility and adaptability, this improved performance can sometimes reduce the generalization and robustness of pre-trained models. To mitigate these issues, this study introduces LePB-SA4RE, a novel ABSA model that integrates the Bidirectional Encoder Representations from Transformers (BERT) architecture with a hard template-based PT method and embeds a lexicon-enhanced dynamic modulation layer. Specifically, the activation function of this layer incorporates weights designed with sentiment-oriented dynamic parameters to enhance the sensitivity of the model to diverse sentiment inputs, and a sentiment lexicon containing three hundred thousand word–sentiment polarity pairs is embedded into the model as additional semantic cues to increase prediction accuracy. The model retains the stability benefits of Hard-prompt methods while increasing the flexibility and adaptability necessary for ABSA in requirements elicitation from app reviews. Experimental results indicate that the proposed method surpasses state-of-the-art methods on the benchmark datasets, and the generalization of the model achieved the highest relative improvements of 72% and 36.6% under low-resource data settings and simulated noisy conditions. These promising findings suggest that LePB-SA4RE has the potential to provide an effective requirements elicitation solution for user-centric software evolution and maintenance.