Towards a mountain biodiversity knowledge graph

Ongoing changes in mountain biodiversity have important consequences for the future provision of ecosystem services across scales and for human livelihoods and wellbeing worldwide, calling for effective action. However, the formulation of environmental policies and measures that address the challenges of sustainable management and conservation of mountain ecosystems relies on knowledge that is ‘trapped’ inside a vast and rapidly increasing corpus of unstructured text - the scientific literature, which to date is not accessible to machine-based approaches. Our objective is to develop MoBiKo, an open access global mountain biodiversity knowledge graph built from the entities and relations extracted from the corpus of mountain biodiversity literature. This knowledge graph will ‘liberate’ and structure available knowledge pertaining to the state of, trends in, and drivers of mountain biodiversity. By following principles of findability, accessibility, interoperability, and reusability, we enable broad usage, its expansion with new entities of interest, and its application for varied downstream tasks. Here, we present ongoing work towards achieving a first version of MoBiKo with (i) an approach to improve named-entity recognition based on a hybrid framework that combines structured resources with large language models, and (ii) a preliminary attempt towards relationship extraction using models that are pre-trained on existing datasets and fine-tuned on synthetically generated mountain biodiversity triplets. In addition, we present the domain-specific gazetteers used to address widespread issues of heterogeneous terminologies and enable targeted inference and efficient pre-filtering of relevant sentences, and we provide examples of the contribution of such gazetteers to linked open data and to the systematic mapping of mountain biodiversity literature. Our preliminary results highlight the potential of hybrid and iterative natural language processing pipelines to bridge rule-based and generative methods. By developing this structured, curated, and digitally accessible knowledge base, we aim to support scientific research and inform policy as well as conservation efforts. We further contribute to “opening up what is known about biodiversity” and thereby support the Disentis Roadmap 2024 vision to fully “leverage the power of biodiversity knowledge from research publications within an open science framework”.