Hengxing Zou

@hengxingzou.bsky.social

Postdoc at Utah State University and Emory University. Species interactions under global change. Theory, data, birds. Previously Institute for Global Change Biology, UMich. PhD Rice University. UChicago alum. Made in Beijing.
hengxingzou.github.io

@hengxingzou.bsky.social

The prompts used are different. The first asks for yes/no interaction, the second sorts them into categories, one of which is “no interaction”. Maybe with more options than yes/no, FP rate is lower because it doesn’t count when a comment identified as another interaction but not a “no interaction”

@hengxingzou.bsky.social

Possibly, but just to clarify, these are different tasks—the Texas comments were used to test if there *are* interactions in the comment, while the Michigan comments were used to categorize the actual interactions, so they can’t be directly compared.

@hengxingzou.bsky.social

Thanks! The 485 comments were randomly selected from the set of comments with interactions. Out of the 485 comments, 5 were false positives (LLM-interaction, manual-no interaction), and 9 were false negatives (LLM-no interaction, manual-interaction). Hope this helps!

@annttate.bsky.socialReposted by @hengxingzou.bsky.social

Please RT: Do you use math to address interesting biological problems? Join us at Vanderbilt BioSci! We are conducting an open-rank TT search in Theoretical Biology: apply.interfolio.com/194619.
We are a friendly and collaborative department with interests spanning a range of biological disciplines!

@hengxingzou.bsky.social

So if you use participatory science platforms and see two birds fighting or a bee visiting a flower, write it down! You could be collecting crucial interaction data for ecologists.

@hengxingzou.bsky.social

Personally, I think this is a very “rustic” use of LLMs: they were designed to parse natural languages. This data collection process also echoes the tradition of ecology that builds on natural history and observations, only supercharged by advanced technology.

@hengxingzou.bsky.social

What can these datasets do? With the broad coverage and exact time/location info, we can examine turnover of species interactions across latitudes, altitudes, urban-rural gradients, and land use changes. They can even help to uncover previously unrecorded interactions!

@hengxingzou.bsky.social

With the help of large language models, we can extract information like this from large amounts of observation notes around the world, constructing datasets with broad spatial and temporal coverage previously unimaginable.

@hengxingzou.bsky.social

These observation comments (here’s one of mine on eBird) can contain information on species interactions (like this Red-tailed Hawk being mobbed by the Red-winged Blackbird), but historically they have been underutilized because they are not in any kind of standardized format.

An eBird comment for Red-tailed Hawk that reads, One with a squirrel and was mobbed by a red-winged blackbird
@afredston.bsky.socialReposted by @hengxingzou.bsky.social

good morning. I strongly suspect that we don't really know (or agree) what geographic ranges are, and therefore, our estimates of how species' ranges are shifting in the Anthropocene are uninterpretable.

I wrote something about this: doi.org/10.1111/2041...

and I'll share the main findings...

Older posts