Improving biomedical entity linking for complex entity mentions with LLM-based text simplification.

Area of Science:

• Biomedical Informatics
• Natural Language Processing

Background:

• Biomedical research and healthcare generate vast amounts of free-text data.
• Entity linking is crucial for accessing this information via natural language processing (NLP).
• Complex, multi-token entity mentions pose challenges for accurate normalization and concept mapping.

Purpose of the Study:

• To develop a method for preprocessing complex entity mentions in biomedical text.
• To improve candidate generation for entity linking using text simplification.
• To evaluate the approach on the BioCreative VIII SympTEMIST shared task.

Main Methods:

• Utilizing generative large language models for text simplification of complex entity mentions.
• Applying few-shot prompting with a Generative Pre-trained Transformer (GPT) model.
• Integrating the simplified mentions into an entity linking pipeline for candidate generation and reranking.

Main Results:

• Text simplification resulted in more easily normalizable mention spans.
• Recall during candidate generation improved by 2.9 percentage points.
• Top-1 accuracy was enhanced by translating recall improvements into reranking, achieving 63.6% on the test set.

Conclusions:

• Generative large language models can effectively simplify complex biomedical entities for improved NLP.
• The proposed text simplification method enhances entity linking performance, particularly recall and accuracy.
• The approach is integrated into the open-source xMEN toolkit for broader application.