Benchmark for Evaluating Knowledge Graph Construction and GNNs
A new benchmark jointly evaluates Graph Neural Networks (GNNs) on noisy, text-derived knowledge graphs and the effectiveness of graph construction methods. Built in the biomedical domain from a single textual corpus, it includes two automatically constructed graphs from different extraction methods and a high-quality expert-curated reference graph as an upper performance bound. The benchmark addresses the challenge of assessing whether performance stems from the learning model or graph quality.
Key facts
- arXiv:2605.05476
- Announce Type: cross
- Knowledge graphs automatically constructed from text are used in real-world applications
- Noise, fragmentation, and semantic inconsistencies affect GNN performance
- Benchmark is dual-purpose: evaluates GNNs and graph construction methods
- Built in the biomedical domain from a single textual corpus
- Includes two automatically constructed graphs and one expert-curated reference graph
- Reference graph serves as an upper performance bound
Entities
—