Paper Type

Complete

Paper Number

PACIS2026-1786

Description

Biomedical knowledge graphs (KGs) have become fundamental infrastructures for biomedical research and clinical decision support. However, large-scale biomedical KGs are predominantly constructed through automated extraction from scientific literature, which often results in incomplete, outdated, or noisy relational information. Furthermore, many meaningful associations may not appear explicitly as extracted triplets due to extraction limitations. To address this limitation, we propose a New Relation Classification (NRC) framework for predicting unseen relations in biomedical knowledge graphs. The proposed method consists of two stages: (1) a knowledge graph embedding (KGE) training stage, and (2) a triplet classification training stage to determine whether a biomedical triplet forms a valid specific new relation. Using SemMedDB and 879 domain-expert-annotated biomedical triplets, we evaluate the proposed method. Our findings suggest that the proposed NRC framework provides a promising direction for dynamically extending biomedical knowledge graphs and improving the reliability of downstream biomedical applications.

Comments

14-Healthcare

Share

COinS
 
Jul 5th, 12:00 AM

From Known Relations to New One: Learning to Classify Triplets of Unseen Relations from A Biomedical Knowledge Graph

Biomedical knowledge graphs (KGs) have become fundamental infrastructures for biomedical research and clinical decision support. However, large-scale biomedical KGs are predominantly constructed through automated extraction from scientific literature, which often results in incomplete, outdated, or noisy relational information. Furthermore, many meaningful associations may not appear explicitly as extracted triplets due to extraction limitations. To address this limitation, we propose a New Relation Classification (NRC) framework for predicting unseen relations in biomedical knowledge graphs. The proposed method consists of two stages: (1) a knowledge graph embedding (KGE) training stage, and (2) a triplet classification training stage to determine whether a biomedical triplet forms a valid specific new relation. Using SemMedDB and 879 domain-expert-annotated biomedical triplets, we evaluate the proposed method. Our findings suggest that the proposed NRC framework provides a promising direction for dynamically extending biomedical knowledge graphs and improving the reliability of downstream biomedical applications.