The exponential growth of user-generated content on Chinese social platforms introduces three critical challenges for Named Entity Recognition (NER): (1) context-dependent ambiguities in informal communication, (2) structural complexity from nested entities, and (3) performance degradation under noisy text conditions. To address these challenges, we propose BERT-CHAB, a unified framework combining hierarchical context modeling with adaptive boundary detection. Our architecture employs two novel mechanisms: dynamic gating for multi-scale feature integration and noise-sensitive attention for boundary confidence calibration. Evaluations across diverse text domains demonstrate state-of-the-art performance: 76.7% F1 on social media analysis (WeiboNER), 95.9% on formal news parsing (MSRA), and 93.75% on nested entity recognition (People’ Daily). The model’s effectiveness stems from its dual capability: the hierarchical component resolves nested entity ambiguities through document-level coherence modeling, while the adaptive attention mechanism preserves syntactic patterns under noisy conditions. Notably, the framework maintains consistent performance in low-resource scenarios without domain-specific optimization, achieving <2% performance variance across data scales. This balanced architecture provides a robust solution for Chinese NER tasks spanning formal documents and colloquial social media content.