HiPrompt: Few-Shot Biomedical Knowledge Fusion via Hierarchy-Oriented Prompting.

Area of Science:

• Biomedical Informatics
• Artificial Intelligence in Medicine
• Knowledge Representation

Background:

• Comprehensive biomedical knowledge bases are crucial for enhancing medical decision-making.
• Knowledge fusion requires integrating diverse knowledge graphs using uniform index systems, often with hierarchical organization for fine-grained granularity.
• Existing unsupervised methods for biomedical knowledge fusion (BKF) lack semantic richness due to reliance on lexical and structural matching.

Purpose of the Study:

• To address the challenge of scarce supervision in biomedical knowledge fusion (BKF).
• To bridge the gap between scarce-labeled BKF data and the data requirements of neural embedding models.
• To propose a supervision-efficient framework that utilizes the few-shot reasoning capabilities of large language models.

Main Methods:

• Developed HiPrompt, a novel framework for knowledge fusion.
• Employed hierarchy-oriented prompts to elicit few-shot reasoning from large language models.
• Utilized a supervision-efficient approach to overcome data scarcity in BKF.

Main Results:

• Demonstrated the effectiveness of HiPrompt on the KG-Hi-BKF benchmark datasets.
• Showcased the ability of HiPrompt to capture rich semantics of biomedical entities and terms.
• Validated the framework's performance in a supervision-efficient manner.

Conclusions:

• HiPrompt offers an effective solution for biomedical knowledge fusion.
• The framework successfully integrates knowledge graphs by leveraging semantic understanding and hierarchical context.
• HiPrompt represents a significant advancement in applying large language models to complex biomedical informatics tasks.