Another weekend drop: Map of human-wildlife conflicts reported in the media

[Aaditeshwar Seth]

Hi all,

GDELT which I had mentioned earlier is Google's indexing platform for news articles. It is very much alive and Cursor was able to use it to get a list of probable news reports of human-wildlife conflicts, deploy LLMs to determine more precisely whether this indeed was a conflict event resulting in deaths or injuries of humans or animals, and use LLMs to extract the most precise location as possible. Check out https://act4d.iitd.ac.in/gdelt-wildlife/. As before, users can suggest edits and moderators can review the edits which get commited to git. 

And this is a generic setup. Anyone can write their own meta json to build similar maps for other kinds of events. I'm hoping to build one next for crop damage events like due to droughts, floods, waterlogging, high temperature, pests, etc. Claude even gave a meta prompt to build the meta json! Check out the README here: https://github.com/aaditeshwar/gdelt-wildlife/tree/main/meta

Repo: https://github.com/aaditeshwar/gdelt-wildlife/

Overall learning: A reasonable GPU workstation running local LLMs is able to parse news articles, documents, etc. to get reasonably well structured outputs. Some of our students are trying to use this to parse research papers and extract variables and relationships to build knowledge graphs that can provide an explainable and disciplined reasoning framework for questions asked by users. 

Slippery slope: The speed at which all this can be done is amazing and exciting but verification and fine tuning seems slow and hard work in comparison. I didn't bother checking for example whether Claude produced a good list of categories, and I didn't check beyond half a dozen cases about how correctly the locations were being extracted or conflicts were being identified. In the agromet-advisory, I had used the LLMs in reverse to cross-check the outputs, and similarly maybe one way is to use a different LLM to cross-check the outputs. But overall, there is a serious need to teach us patience and discipline, that's the only way I feel. Principles like supplying the provenance, allowing humans-in-the-loop, explaining the setup, etc. can easily be done as a compliance measure and will not add much value I think. 

Adi

--

Aaditeshwar Seth
Microsoft Chair Professor, Computer Science and Engineering, IIT Delhi

Co-founder, Gram Vaani; Co-founder, CoRE Stack

Research | CoRE Stack | Gram Vaani | ACT4D 

Technology and (Dis)Empowerment: A Call to Technologists

Latest@GV: Facebook | Twitter | LinkedIn

[Aaditeshwar Seth]

Hi all, 

I realized that we had hit some rate limits earlier and have updated the map for human-wildlife conflicts: https://act4d.iitd.ac.in/gdelt-wildlife/. Have also crawled for events related to crop damages and for avian flu and wild bird deaths. The crops data is not fully updated, should get done by tomorrow. A dashboard is also added. 

Cursor Composer 2 did 99.9999% of the work. Coding LLMs have started offering more usage for weekends! 

There is quite a lot of room for improvement like some invalid events are still getting through, some events have multiple reports that didn't get de-duplicated, and some thought can be put to flag potential inaccuracies that can be manually reported. 

I was thinking that we could use reports of pests attacks to build a predictive model trained on rainfall, temperature, etc. data of the location, which when fed weather forecast data will be able to forecast pest risk. It's like species distribution modeling where we only have positive samples and need to create pseudo absences for the ML model. But so far there are too few pest incidences being reported. 

Adi