Machine learning to identify good practices for climate action with Nature-based Solutions (KB-46-005-018)

We created a corpus of unstructured content representative for the topic of NbS. As a next step in the research, we explored the definition of the domain and its semantics using a top-down as well as a bottom-up approach. We explored a combination of technologies from the realm of Natural Language Processing (NLP), like a PDF parser, transformers, like ClimateBert, and named entity recognition (NER). By doing so, we aimed to extract and classify NbS as well as their associated barriers and enablers for implementation from text, and store this in a database.

The codebook, and particularly specific fragments describing targeted categories, has been used in initial experiments to find and analyse scientific publications. During these explorations it became apparent that the terminology surrounding NbS is rather unformalised and that the boundaries between taxa can be unclear. As an alternative to the above approach, we explored Google’s Bard generative AI to propose a taxonomy for NbS, thereby effectively leveraging contextual information on NbS in Bard’s training data. The result shows resemblance to the codebook manually conceived but appears to follow more strictly defined taxa. Understanding the location of the NbS was a challenge. Among the implemented models, ChatGPT consistently yields the most accurate results. To recognise terms signalling barriers and enablers, our first crude attempt was a rule-based tagger developed in the Spacy framework.

Lessons learned:

• The language around NbS is still too fuzzy for AI to properly analyse it.

• ChatGPT is the most promising tool but has limitations regarding data security, reliability and traceability of results.

• Colleagues working on AI in different science groups know how to find each other. Non-AI colleagues know who can help them with these methods.

• The generated taxonomies for NbS look interesting.