SiBERT: A Siamese-based BERT network for Chinese medical entities alignment.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Entity alignment is crucial for integrating medical knowledge graphs.
• Current deep learning methods lack term-level embeddings, limiting performance and increasing computational costs.

Purpose of the Study:

• To propose a Siamese-based BERT (SiBERT) model for improved Chinese medical entity alignment.
• To address limitations in term-level representation and computational overhead in existing methods.

Main Methods:

• Developed SiBERT, a Siamese-based BERT model generating term-level embeddings from word embedding sequences.
• Pre-trained SiBERT with a public synonym dictionary and fine-tuned it on labeled medical entities (disease, symptom, treatment, examination).
• Utilized an entity alignment algorithm to select the most similar standard term.

Main Results:

• SiBERT demonstrated superior alignment accuracy compared to existing algorithms.
• The proposed method significantly improved computational efficiency.
• Extensive experiments on real-world datasets validated the effectiveness of SiBERT.

Conclusions:

• SiBERT offers a significant advancement in Chinese medical entity alignment.
• The model's term-level embedding approach enhances feature representation for similarity calculations.
• SiBERT provides a more accurate and computationally efficient solution for medical knowledge graph integration.