Graduate Researcher Co-Authors Environmental Evidence Study
Check out this recent article published in Environmental Evidence. Co-authored by M.Sc. Geography candidate Steve Robinson with colleagues from the Department of Biology and collaborators from the University of Waterloo and the University of California.
In this artificial intelligence (AI) methodology paper, a student-lead team trials and validates the use of OntoGPT, a tool for extracting information from text with large language models (LLMs) and ontology-based grounding, to assist data extraction for ecological evidence synthesis. Ontologies—expert-curated knowledge bases containing information about a domain of interest, including concepts and their relationships—are designed to be human- and machine-readable, and are a foundational part of the Semantic Web. Ontology grounding has the potential to reduce hallucinations (false information) generated by LLMs, thereby increasing the reliability of AI-assisted information extraction.
The authors compared information extracted from a sample of coastal wetland restoration literature by manual reviewers and OntoGPT and found moderate overall agreement. Importantly, hallucinations were very rare in the OntoGPT data. The results suggest that ontology grounding can improve the veracity of AI-extracted information, but that ecological ontologies are not yet sufficiently developed to provide the broad and nuanced conceptual coverage needed for thorough, autonomous AI-driven reviews.