Description
Ecological synthesis is increasingly difficult not only due to the growing volume of ecological literature, but also because evidence is deeply contextual. A claim about range shifts or niche change depends on the taxon, habitat, method, driver, or response under consideration. Standard retrieval-augmented generation can retrieve relevant text but often leaves this structure implicit. We present EcoGraphRAG, a pilot workflow that represents ecological claims as causal evidence records and queries them through a provenance-preserving graph. We applied EcoGraphRAG to a 58-paper corpus of ecological studies on niche breadth, range shifts, specialization, and climate-related responses. Full-text sections were parsed, and we extracted evidence records with a task-specific causal evidence schema and validated records before graph construction. The corpus contained 296 validated records and a graph with 1,883 nodes and 3,126 edges linking papers, claims, passages, taxa, drivers, responses, methods, and concept mappings. We compared graph retrieval with a semantic retrieval baseline across synthesis questions using the same evidence corpus. Retrieved evidence was evaluated for relevance, passage support, ecological context, and concept mapping. Both approaches returned well-supported evidence, but the graph representation made driver-response structure, provenance, and concept coverage easier to inspect. A provenance-matching check recovered 289 of 296 extracted source texts from parsed papers, while flagging a small number of page-attribution and PDF parsing issues for manual review. Our work suggests that, rather than automating synthesis, EcoGraphRAG makes the evidence behind synthesis easier to inspect. By keeping claims connected to ecological context and source text, causal evidence graphs can support more transparent ecological synthesis when provenance, taxa, methods, and drivers matter as much as textual similarity.
| Topic | Topic 4: Can knowledge graphs and AI-based synthesis improve ecological synthesis? |
|---|