This paper provides a review of the application of Bidirectional Encoder Representations from Transformers (BERT) in the field of Requirements Engineering (RE), a domain where the accurate interpretation of natural language requirements is crucial and challenging. Unlike existing reviews that broadly cover Natural Language Processing (NLP) techniques in RE, this study focuses specifically on BERT and its derivatives, offering a fine-grained taxonomy of models, tasks, and optimization strategies. The novelty of this work is in systematically bridging advances in Transformer-based language modeling with the specific demands of RE. BERT, with its dual context understanding and transfer learning capabilities, has made significant advances in tasks such as requirement classification, information extraction, and requirements management. By synthesizing recent studies from 2020 to 2024, we identify the dominant trends, categorize models into direct usage, domain-specific fine-tuning, and hybrid architectures, and provide statistical analyses of datasets and model adoption. In addition, we highlight unresolved challenges, such as data scarcity, interpretability, and computational efficiency, and propose future research directions. This work contributes both a structured roadmap and practical guidance for researchers and practitioners seeking to leverage BERT in complex RE tasks.