Where graphs supplement LLMs
Graph-based parsers appear to outperform LLMs on relation extraction — and the gap widens as relational complexity grows. A preprint out today from Gajo et al. has evidence across six datasets. For pharma and biomedical knowledge graphs, where the useful relations are mechanism-of-action chains and adverse event pathways rather than simple co-mentions, this is the relevant regime. Useful alongside what I wrote earlier this week on knowledge graphs as research discovery tools.