Medical Entity Recognition Based on Self-Attention and Adversarial Training

In recent years, entity recognition methods have faced numerous challenges, including large parameter sizes, low computational efficiency, interference from label noise, and incomplete semantic information. To address these issues, we propose a model architecture based on self-attention mechanism and adversarial training. The model effectively leverages ALBERT's embeddings and glyph features for enhanced representation, and utilizes adversarial training to improve robustness. The dual-layer self-attention mechanism captures dependencies both within and between sentences. Experimental results demonstrate the effectiveness of our model.