Brief description of the organization
EcoWeaver is an organization developing decision support tools for ecological restoration and conservation practitioners. We have created an open knowledge framework linking causal ecological evidence to restoration practice, built on the Causal Mosaic Schema (CAMO) — a schema co-developed with philosophers of causality for documenting not just that X causes Y, but how, and how confidently. It's developed out of a postdoctoral translational-ecology research program spanning the Bennett Lab (Carleton University) and DRAGEN Lab (University of Waterloo).
Problem area
Ecological literature contains rich causal claims about restoration and conservation interventions and their outcomes, but that knowledge is locked in unstructured prose. Manual extraction doesn't scale, and generic LLM extraction without a rigorous causal schema tends to flatten important nuance — correlation vs. causation, mechanism, confidence — into oversimplified graphs.
Main objectives
- Build and stress-test a pipeline that turns a corpus of open-access restoration-ecology papers — starting with a 308-paper set drawn from FACETS (Canada's multidisciplinary open-access science journal) — into structured causal claims.
- Extract those claims using the CAMO (Causal Mosaic Ontology) schema, a LinkML-based schema co-developed with philosophers of causality, which documents not just "X causes Y" but the mechanism and confidence behind the claim.
- Evaluate extraction quality — accuracy, richness, scalability — against a held-out, hand-annotated benchmark set.
Scope of work
- Review of the CAMO schema and worked examples
- Corpus ingestion and preprocessing of the FACETS 308-paper set (extensible to a larger corpus)
- Design and testing of LLM-based extraction methods (prompting, fine-tuning, or a hybrid pipeline)
- Evaluation against a hand-annotated gold-standard set
- Documentation and open-source release
Deliverables
- Report
- Website
- New protocols/processes
- Resources
- Presentation
- Reproducible, re-runnable pipeline for knowledge extraction. A delegate from the University of Waterloo will have the opportunity to present at an international conference in Germany in March 2027. (potential for travel subsidy; or virtual)
Team meeting frequency
Weekly to start, shifting to bi-weekly as they find a rhythm.
Skills and training required
- Python
- NLP/LLM prompting or fine-tuning
- Familiarity with knowledge graphs or ontologies (LinkML a plus)
- Comfort reading primary ecological literature (trainable)
- An interest in the philosophy of causality is a bonus, not a requirement.
Resources required
- Access to the Nibi supercomputer and/or GPU workstations for LLM inference (some provided)
- The FACETS corpus (open access) and CAMO starter repo (provided)
- A modest LLM API budget if not run fully locally.
NDA or a commercialization agreement for this project?
Yes