Do Syntactic Features Help Biomedical Relation Extraction? An Empirical Study of Verb Token and Dependency Graph Augmentation
Year: 2026
Type: conference-paper
Abstract: We investigate whether explicit syntactic features improve transformer-based biomedical relation extraction when added to typed entity marker pooling.We evaluate two augmentation strategies on top of BiomedBERT: (1) verb token augmentation, which concatenates the hidden state of the dependency root verb to the entity representations, and (2) a two-layer graph convolutional network (GCN) that refines encoder hidden states over the dependency parse before entity pooling.We experimented on three biomedical datasets: ChemProt, DDI, and AIMed with three random seeds.We found neither strategy consistently outperformed the entity-only baseline.The GCN yielded modest gains on AIMed (+0.007F1) and ChemProt (+0.003F1) but decreased performance on DDI (-0.013F1).Verb token augmentation helps only on AIMed (+0.004F1) and underperforms on the other two datasets.A syntactic characterization of the datasets reveals that DDI has substantially higher passive voice usage (50.7% of relation-bearing sentences) than AIMed (27.0%) or ChemProt (30.9%), suggesting that syntactic augmentation is more effective when sentences exhibit active verbal structure with semantically informative predicates.These results suggest that corpus-level syntactic characteristics, particularly passive voice usage, may moderate the utility of explicit syntactic augmentation, though the small magnitude of observed differences warrants caution in interpretation.
Authors: Mustafa Sikder, Ernest Kwegyir-Afful
Institution: United States Food and Drug Administration
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
OpenAlex ID: W7166825549