<![CDATA[Biomedical Named Entity Recognition (NER) plays an important role in Clinical Natural Language Processing by enabling the automatic extraction of disease-related entities from medical text. This study compares the performance and stability of three sequence labeling approaches: Conditional Random Field (CRF), Bidirectional Long Short-Term Memory (BiLSTM), and BiLSTM-CRF. Experiments were conducted using the NCBI Disease Corpus, comprising 5,392 training instances, 918 validation instances, and 937 test instances, with a class imbalance ratio of 11.06:1. Each model employed 200-dimensional GloVe word embeddings, which achieved 78.2% vocabulary coverage. To assess the consistency of model performance, experiments were repeated across five different random seed configurations. The results demonstrate that BiLSTM-CRF achieved the highest F1-score of 78.86% (±0.34%), outperforming CRF at 78.49% and BiLSTM at 76.73% (±0.86%). Compared with BiLSTM, the hybrid model improved the F1-score by 2.13 percentage points, while its improvement over CRF reached 0.38 percentage points. BiLSTM-CRF also exhibited the most stable performance, as reflected by its lower standard deviation. Learning curve analysis further showed that BiLSTM-CRF consistently surpassed the other approaches after utilizing 40% of the training data. These findings suggest that combining BiLSTM's contextual representation capability with CRF-based sequence dependency modeling provides a more effective and stable approach for disease entity recognition in biomedical text..]]>
Copyrights © 2026
