In software development, classifying software requirements is essential since it has a direct impact on the overall quality of the system. Accurate classification requires a comprehensive understanding of requirements at multiple levels, including words, sentences, and entire documents, to ensure all nuances are considered. To address these challenges, this paper introduces the Hierarchical Attention Network with a Bidirectional Long Short-Term Memory (H2AN-BiLSTM) model, a hybrid approach that integrates document-level and sentence-level semantics from Doc2Vec and contextualized word-level embeddings from DistilBERT. An unlabeled dataset of 12,495 user reviews from various applications on the Play Store was annotated using a BERT-based uncased model fine-tuned on the publicly available PROMISE dataset. The annotated dataset is subsequently enhanced through a Hierarchical Attention Network (HAN) and BiLSTM. Additionally, the model effectively classified software requirements (SRs), achieving a test accuracy of 94.40%. This research not only advances the state of the art in software requirements classification but also paves the way for research into more sophisticated approaches that leverage hybrid Hierarchical Attention Network (HAN) models to enhance classification performance across diverse domains.