Path-based knowledge reasoning with textual semantic information for medical knowledge graph completion.

Area of Science:

• Artificial Intelligence
• Bioinformatics
• Natural Language Processing

Background:

• Medical knowledge graphs (KGs) are crucial for discovering new medical facts but suffer from significant incompleteness.
• Existing path-based knowledge graph completion (MedKGC) methods struggle with the sparsity of entities and paths in medical KGs.
• Traditional MedKGC approaches overlook the textual semantics within KG paths, limiting performance improvements.

Purpose of the Study:

• To develop novel path-based reasoning methods for medical knowledge graph completion (MedKGC).
• To address and alleviate the issues of entity and path sparsity in medical KGs.
• To integrate textual semantic information into MedKGC for enhanced reasoning capabilities.

Main Methods:

• Proposed two novel path-based reasoning methods specifically designed for MedKGC.
• Utilized the pre-trained BERT model to incorporate textual semantic representations of entities and relationships.
• Modeled symbolic reasoning in medical KGs as a numerical computation problem within textual semantic spaces.

Main Results:

• The proposed methods demonstrated significant improvements over state-of-the-art path-based knowledge graph reasoning techniques.
• Achieved an average performance increase of 5.83% across all relations on a Chinese symptom knowledge graph.
• Successfully addressed entity and path sparsity issues in medical knowledge graph completion.

Conclusions:

• Introduced two new knowledge graph reasoning algorithms that effectively mitigate sparsity problems in MedKGC.
• Pioneered the use of pre-trained language models and text path representations for medical knowledge reasoning.
• Provided an interpretable method for completing impaired symptom knowledge graphs, outperforming existing approaches.