I built an LLM-assisted database of plant name etymologies — is this a useful approach?

Nature Talk
1

I built an LLM-assisted database of plant name etymologies — is this a useful approach?

Hi everyone,

I recently created a web project called Etymon Plantae, a searchable database for exploring the etymology of plant scientific names.

https://lightbox-archive.com/etymon-plantae.html?lang=en

The database currently focuses on plant taxa found in Japan and is intended to cover the Japanese flora comprehensively. You can search by a scientific name, genus name, or specific epithet and see an explanation of the possible origin and meaning of the name.

Although the dataset is centered on Japan, many genera and species are shared with other parts of the world, so I thought it might also be of interest to iNaturalist users outside Japan.

What interests me most about scientific names is how much information can be hidden in them: references to morphology, habitat, geography, mythology, people, historical circumstances, and sometimes rather unexpected stories.

However, there is an important aspect of this project that I would especially like to discuss here.

The etymology descriptions were created with the assistance of an LLM.

I am aware that using an LLM for this kind of information raises legitimate questions about accuracy, hallucination, reliability, sourcing, and whether AI-generated explanations are appropriate at all for a resource dealing with scientific nomenclature.

For that reason, I am not only looking for corrections to individual entries. I would genuinely like to hear critical opinions about the basic idea behind the project.

For example:

  • Do you think using an LLM to investigate and explain the etymology of scientific names is a reasonable approach?

  • Can an LLM-assisted resource like this be useful if its limitations are clearly stated, or does the possibility of fabricated or incorrect etymologies make the approach fundamentally unreliable?

  • What level of citation, sourcing, or human verification would you expect before considering such information useful?

  • Are there certain types of etymological claims that should never be inferred by an LLM without a primary or authoritative source?

  • Would it be better to present several possible interpretations when the origin of a name is uncertain, rather than giving a single explanation?

  • If you think this approach is likely to create more misinformation than useful information, would you recommend that I discontinue the site rather than continue developing it?

  • On the other hand, if you think the concept has value, what changes would make it more responsible or scientifically useful?

  • Would links to original descriptions, botanical literature, dictionaries of Greek and Latin roots, or other references substantially improve the usefulness of the database?

  • If the project were expanded beyond the flora of Japan, are there particular countries, regions, or floras that you think would be worth adding?

  • More broadly, what would make something like this genuinely useful to botanists, naturalists, and iNaturalist users rather than simply an interesting AI experiment?

I am genuinely open to criticism, including criticism of the premise of the project itself.

I do not have a strong commitment to keeping the site online if people with more experience in botanical nomenclature and taxonomy feel that presenting LLM-assisted etymologies in this way is inappropriate or potentially misleading. I would rather hear that criticism directly than continue developing something that experts consider fundamentally unsound.

At the same time, if there is a responsible way to build this kind of resource — for example through better sourcing, uncertainty labels, community corrections, or human verification — I would be very interested in exploring that direction.

If you are curious, please try searching for a scientific name you know well. I would be especially interested in examples where the explanation is clearly wrong, questionable, incomplete, or surprisingly good.

Any criticism, corrections, suggestions, or thoughts about the use of LLMs for this kind of project would be very welcome.

Thanks for reading!

2

etymology is a real field of study. It can be hard because there are a lot of “folk etymologies” which are kind of like just so stories about why words are the way that they are.

With scientific names the first place to look would be the original type specimen description. Sometimes the author will explain why they have chosen a particular name. I wish this were a requirement but it’s not. But if this information exists it would be correct.

I could see using the LLM to generate a list of links and references to the original descriptions, maybe even with the page number, then you could verify that those are correct and those etymologies would be correct.

What about the rest of the binomials that don’t have the answer in the original description? For those you could ask for a different source. I think these sources would be much harder to verify since they would span many topics and the danger of using a poor source grows.

For some of the binomials it will cite Latin roots or other languages, so to verify it you could have someone who knows those languages look at the results. Of course it’s not like Latin or any languages is always used correctly in a binomial.

So, through this process you could save a lot of time, but still make a verified list.

3
  • I like the idea and the clear interface, congrats
  • I need sources! (… and preferably non-hallucinated ones!)
  • Why the 12.5MB download? how would it scale if going beyond Japanese flora?
4

By the way, I love the impulse behind this project. As a teacher I’ve given a few presentations about ants for kids. I tell them about scientific names and why they are useful. Every single time I say a scientific name they always ask “why is it called THAT?”

This is, in truth, my first question too. It’s not always easy to answer. In some cases these reasons are lost to time. Which is why I wish that they were a required and expected part of a type description, not just a “nice to know.”

I think there was this idea that everyone doing science would know Latin and the binomials would always be based on Latin (I’m very glad that this limitation is going away) so it would be “self evident” to the scientist what the name meant. And there are moments where that is true:

Cataglyphis nigripes

Has dark legs.

Lasius flavus

Is fuzzy and yellow.

However, just as often the tree has been rearranged, and the reason can be lost to time. I think we should talk about it a lot more, choosing names should be a more public and well-explained process where possible. That is, if we really want these names with their better precision and consistency to be popular.

I decided to try to use the LLM to do this process to see what would happen. It helpfully pointed me to "Johann Christian Fabricius, Species insectorum, Volume I. Formica flava—the species now called Lasius flavus—appears on page 491" There in the pdf I found a frustrating brief description in Latin. It doesn’t explicitly say “I picked the word flavus because it’s a yellow ant.” The paper is just trying to tack a name on to every ant and doesn’t give them much individual attention. The descriptions are all relative to each other such as “small” which means “smaller than the other ants in Formica” … And that’s another wrinkle Lasius flavus used to be in Formica until the genus grew so bloated and confusing it had to be revised. In Mayr 1861 — Die europäischen Formiciden our little yellow fuzzy ant is moved from Formica to Lasius with many many other ants. At this point I felt that I’d seen enough to support my original idea. It’s named Lasius flavus because it’s yellow.

The LLM did a good job finding the original type description. I’m only really comfortable asking it to do that because if it was wrong I could tell. These systems will just spit out stuff and it will sound true. But then it turns out to be wrong, so if I don’t know how to check it? It is useless information.

I think that could be a big danger with etymologies. It’s easy to make up story about a name that sounds true.

5

Thank you very much for the comments.

I agree that referring to the original description is the best approach whenever possible.

Regarding the 12.5 MB download, the current implementation loads the entire database at once. This could certainly be made much lighter by changing the way the data is queried and loaded.

I have also been developing another application that takes a scientific name, searches for the original description, and uses OCR to extract information from it. If I can run that process in my spare time and connect the results with this project, I think it could help provide much more reliable etymologies.

Of course, as you both know, finding the correct original description is often the difficult part. The literature is scattered across many journals, books, archives, and languages, so this is not always straightforward.

When I am able to verify the relevant literature, I would like to add the source to the results for exact scientific-name searches. Providing separate sources for every genus name and specific epithet may be more difficult, although ideally I would also like to link those explanations to reliable references where possible.

Sources really are important, especially for a project like this!

Thank you again — your comments have given me a clearer idea of how this project could be improved.

6

I usually go to PlantZAfrica pza.sanbi.org for our South African plants
But today when I want to link … Bad Gateway
If the sp is already included, they have a paragraph to explain the derivation of the binomial.
For Chrysocoma coma-aurea the daisy is twice golden – chrys and aurea – in Greek and Latin.
I know that site is written by local botanists (some (still) active on iNat) with info from the original description. Wary of an answer which has not at least been checked by a knowledgeable human.

Formatting is broken there but here is a copypasta

Chrysocoma is derived from the Greek, khrusos = gold, kome = hair, lock, in allusion to the. terminal, golden flower heads. Coma-aureus from the Latin, coma meaning a ‘head’ and aureus meaning ‘yellow’.

7

If only after etymology (rather than an extensive compilation of original descriptions) and as you probably already know or noticed, many specific epithets (vulgaris, minor, officinalis…) are heavily reused across taxa, and have their meanings already compiled somewhere (e.g. here or here): no need to trace original species descriptions. Just use such documents as source.

Once you get all these out of the way, the fun begins: less common, or very niche, or faulty epithets (spelling errors, poor latin, …) + all the genus names, many of which fortunately(?) convey little extra information (e.g. ‘Carex’ meaning ‘sedge’ :grinning_face_with_smiling_eyes:) - just reference some dictionary of classical latin and you’re done.

8

Does this project use iNaturalist data or content in some way? I currently don’t see the connection to iNaturalist.

9

Thank you both — these are very helpful suggestions.

I agree that for common epithets such as vulgaris, minor, or officinalis, reliable botanical or linguistic references may be sufficient, without tracing every original description individually. For less common names, unusual spellings, personal or geographical names, and genus names, the original literature will probably become much more important.

One reason I initially kept the etymology generation largely self-contained within the LLM was that I was unsure how far I should go in automatically scraping and reusing information from external websites. Since the relevant literature and databases are spread across many different services with different conditions, I thought it was safer not to collect their content automatically.

However, your comments have made me realize that keeping everything inside the LLM creates a different and probably more important problem: the user cannot verify where an explanation came from.

So I think a better approach would be to cite or link to reliable sources wherever possible, without necessarily reproducing their content.

The PlantZAfrica example and the references to existing compilations of common epithets are very helpful in that respect. Ideally, I would like the site to distinguish between explanations supported by a source and interpretations that are only inferred.

Sources really are important for something like this. Thank you — I’m learning a lot from this discussion.

10

Thanks for asking.

No, the project does not currently use iNaturalist data or content directly.

I posted it in Nature Talk because I understood the category to also include nature-related topics that are not directly related to iNaturalist, provided that they are framed as a discussion rather than simply promoting an external project.

My intention here was specifically to ask the iNaturalist community — especially people with experience in taxonomy, nomenclature, and botany — for critical feedback on whether using LLMs for scientific-name etymology is appropriate, and what kind of sourcing or verification would be necessary.

If I have misunderstood the scope of the category, though, I’m of course happy to have the topic moved or removed as appropriate.

12

It’s nearly universal now for newly published taxa, but older works often omitted the information.

13

Most literature prior to sometime in the 1920s is available online. (It’s out of copyright, so has been picked up by big digitising projects - Botanicus, Gallica, Google Books, RJBM, …) Most material after around 1990 is available online, but substantial portions are paywalled - I reckon on about 50% being accessible either as open access or on preprint servers. Between those two dates availability is more limited - some journals have put up some or all of their back catalog, but others haven’t.

IPNI will tell you where a name was published, and some of the time links to a copy at BHL. BHL indices archive.org and RJBM as well as its own material, so it’s the best first port of call for an old publication. It has some newer works, but coverage is much thinner.

14

If the user doesn’t know where the information came from, and if the LLM-generated output may not reflect accurate etymology, what is the value of the database? Why will people want to use it knowing that they have no way to know what is correct or incorrect (or the source)? Is this just LLM translation of the Latin? Where is the LLM obtaining it’s information?

I tried Anemone, and very little of what it returns appears to reflect what the actual documented etymology is (beyond wind flower)…but much of it is in Japanese: https://www.etymonline.com/word/anemone

On the website, I don’t see any explanation (in English) that describes how the information is being derived (no mention of LLMs).

15

Thank you — this is very helpful.
I would definitely like to use original descriptions wherever they can reasonably be found. At the same time, one of the difficulties I am running into is scale. For a database covering a large number of names, I don’t think it is realistically possible to trace every single name back to its original publication, especially when the literature is inaccessible, difficult to OCR, or does not actually explain the etymology.

So I am beginning to think that the database should distinguish between different levels of evidence rather than treating every entry equally — for example, etymologies confirmed from the original publication, those supported by reliable secondary sources, and those that are linguistic interpretations or otherwise inferred.

That would allow the database to remain reasonably broad while making it much clearer to users which explanations have actually been verified against a source.

Thank you!

16

Thank you — this is a very fair point, and the missing explanation about the use of LLMs was actually my mistake.

The site did previously mention that the etymology descriptions were generated with LLM assistance, but that information seems to have disappeared during a recent update. I have now restored it and added more detailed information about the dataset, its limitations, and confidence levels.

The current descriptions were generated using Claude Opus 5 and ChatGPT 5.6 Sol, based on information already contained in the models rather than by scraping or searching external websites. Because of that, most individual entries do not yet have a traceable source attached to each statement, which is an important limitation.

The Anemone example illustrates this quite well. My database gives a broadly similar linguistic explanation, but it does not include much of the finer historical context that can be recovered from the literature. That is exactly the kind of information that is difficult to capture comprehensively at this scale.

At the same time, one of the goals of the project is breadth and accessibility. The current database covers more than 17,000 scientific names, and I would like users to be able to get a concise indication of a name’s likely meaning very quickly, rather than requiring a literature search for every query.

So I do not see the database as a replacement for a source-based etymological reference. I see it more as a broad first-pass reference: useful for quickly understanding the likely linguistic meaning of a name, while making it clear when the historical reason for a particular name has not actually been verified.
Where original publications or reliable sources can be identified, I would like to add those progressively without losing the broad coverage and simplicity of the current site.

The discussion here has been very useful in clarifying that distinction.

17

If I use this prompt with Lumo AI and Google AI, I get a very detailed narrative that includes all the hypotheses mentioned in the link above, with links to numerous sources.

Prompt–>What is the etymology of the plant genus Anemone?

18

Thank you for your feedback.

Based partly on the feedback in this discussion, I have just added a new section to the search results called “Investigate further with AI.” It can open ChatGPT or Gemini, or generate a prompt that includes the dictionary entry and asks the AI to verify it and investigate the etymology in more detail.

So rather than treating the database and AI search as alternatives, I think a useful workflow may be:

Etymon Plantae → quick overview
AI search → deeper investigation and alternative hypotheses
original literature → verification where necessary

I think this also makes the limitations of the database more useful and transparent: it provides broad coverage and a simple starting point, while giving users an easy way to go beyond the short description when they want to.