BiGran-NER

# 🔍 BiGran-NER: Semantic Alignment with Bi-Granularity for Noisy Clinical NER ## 📘 Overview **BiGran-NER** is a deep learning framework that enhances named entity recognition (NER) performance in noisy clinical texts, such as those found in electronic health records (EHRs), medical reports, and user-generated health content. The model introduces a **Bi-Granularity Semantic Alignment** strategy that synergizes word-level and sentence-level representations to mitigate token fragmentation, out-of-vocabulary challenges, and contextual ambiguity. > 🔬 Proposed in: *"Semantic Alignment with Bi-Granularity for Enhanced Named Entity Recognition in Noisy Clinical Texts" (BMC, 2025)* --- ## ✨ Key Features - **Bi-Granularity Architecture**: - Combines token-level and sentence-level embeddings - Aligns semantic and syntactic spaces via contrastive and dynamic distillation - **Dual Alignment Strategies**: - *Semantic Contrastive Learning (SCL)* to align NER token spans across granularity levels - *Dynamic Knowledge Distillation (DKD)* to transfer uncertainty-aware knowledge from sentence to token predictions - **Noise-Robust Training**: - Tailored for noisy, informal clinical corpora - Efficient adaptation to short-form or fragmented expressions --- ## 🧠 Model Architecture