Automating Biomedical Knowledge Graph Construction For Context-Aware Scientific Inference

AutoBioKG constructs highly accurate, context-aware biomedical knowledge graphs, outperforming baselines by 3.6-17.8 percentage points in zero-shot F1 across multiple datasets.

• Why it matters: Capturing the dynamic and context-dependent nature of biomedical interactions is crucial for accurate scientific inference, yet existing methods oversimplify these complex mechanisms, leading to semantic loss and contradictions.
• What they did: AutoBioKG employs an end-to-end framework using composite triples, trained on BioOpenIE and refined with self-training on unlabeled literature, to encode environmental and entity attributes alongside core relationships.
• The result: The approach enables superior performance in biomedical question answering and knowledge extraction, supporting scalable transformation of unstructured literature into structured, context-aware knowledge graphs.