Nuestros datos se publican bajo varias licencias Creative Commons.Más información
Si le encanta el contenido de este sitio, por favor considere hacer una donación.
Muro (7.137 hilos de mensajes)
Consejos
Antes de preguntar, asegúrese de leer FAQ.
Nuestro objetivo es mantener una atmósfera saludable para las discusiones civilizadas. Por favor, lea nuestras reglas contra el mal comportamiento.
deniko
hace 19 horas
deniko
hace 21 horas
frpzzd
hace 22 horas
araneo
ayer
deniko
ayer
deniko
ayer
deniko
ayer
deniko
ayer
deniko
ayer
mraz
hace 7 días
I already played with the concept I called “Tatoeba clouds” back in 2020 — here’s my original post:
https://tatoeba.org/en/wall/sho...#message_34963
A Tatoeba cloud is all the sentences that are linked with each other, directly or indirectly — meaning that any two sentences in the same cloud are connected through a chain of translations.
In 2020, the biggest cloud contained about 400,000 sentences. Now this same cloud has almost doubled in size.
This time, compared to my approach five years ago, I decided to eliminate links through Toki Pona and Lojban, since they create some unnecessary linking.
The top 10 clouds are these:
https://i.imgur.com/xu22lsh.png
The first cloud, which I call the Infinite Translation Loop, contains about 770,000 sentences. If you also include links that pass through Toki Pona or Lojban, it grows to about 832,000 sentences — so those languages don’t blow the number out of proportion, at least for the first cloud.
I've uploaded the full first cloud here:
https://docs.google.com/spreads...it?usp=sharing
You can filter it by language if you like (for example, only show sentences in English).
For this, select Data-Change View->English Only
You can create other views, I've only created one for English.
Remember, all these sentences are related. So if you see two that seem completely unrelated, like:
13282688 Where did you find out about this?
13282917 Damn, it's cold.
There's still a chain of translations linking them together.
I wrote a script to trace the chain between any two sentences. Unfortunately, it won't work dynamically inside Google Sheets, but if you're curious, give me two sentence IDs and I can send you the chain of translations.
For example, the link between the two sentences above is:
13282688 eng Where did you find out about this?
6919783 lit Kur tu apie tai sužinojai?
1917863 epo Kie vi eksciis pri tio?
1724296 fra Où avez-vous appris ça ?
1723847 eng Where did you learn this?
8258274 ber Amek ay tlemded aya?
4888523 eng How did you learn that?
3980162 tur Onu nasıl öğrendin?
3738777 eng How did you find that out?
8258907 ber Amek ay d-tufid aya?
7813136 eng How did you find this?
3015806 rus Как тебе это?
5660228 deu Wie findest du das?
402404 nld Wat denk je ervan?
937375 deu Was meinst du dazu?
5377218 hun Mit szólsz hozzá?
56368 eng How about that!
219052 jpn こりゃすごい！
1849064 eng This is awesome.
2013320 tur Bu harika.
3734407 eng It's wonderful.
5935872 ukr Це чудово.
2123609 eng It's cool.
1250017 tur Hava serin.
386382 epo Estas iom malvarme.
63615 eng It's pretty cold.
551461 epo Estas tre malvarme.
269824 eng It's bitter cold.
856613 tur Hava çok şiddetli soğuk.
856079 eng It's so fucking cold!
856078 spa Joder, qué frío hace.
13282917 eng Damn, it's cold.
Analysing a cloud like this could help spot wrong links or bad translations — but honestly, I feel like most of these links are perfectly legitimate.
It's like we're all playing a massive, real-life game of Chinese Whispers. A sentence starts in one language, gets translated to another, then another, and over time the meaning drifts. Sometimes it's just tiny shifts — a word choice here, a change of tone there — and sometimes it takes a wild turn. You might begin with “Where did you find out about this?” and, many hops later, arrive at “Damn, it's cold.”
What's amazing is that every step along the way makes sense to the person translating it. Each link is logical in its own context — yet the long chain as a whole can be unrecognisable from where it started. It's a living demonstration of how language, culture, and interpretation can subtly (or dramatically) reshape meaning.
Another pair in the same cloud I randomly picked: “Rome wasn't built in a day.” → “Can you do that?”
It's like starting a conversation about ancient city planning and ending up asking someone to fix your sink. That's the magic of indirect translations: the meaning can drift, twist, and re-emerge in entirely new forms, all through legitimate (I hope!) steps.
The actual chain is:
1777 eng Rome wasn't built in a day.
649178 epo Ne en unu tago elkreskis Kartago.
3535350 tur Sabreden derviş muradına ermiş.
23931 eng Everything comes to those who wait.
137924 jpn 待てば海路の日和あり。
129140 fra Après une tempête vient le calme.
859496 spa Después de una tormenta, viene la calma.
21263 eng After a storm comes a calm.
459360 deu Nach dem Regen kommt Sonnenschein.
36678 eng Every cloud has a silver lining.
1020173 tur Her işte bir hayır vardır.
2891545 eng It's all for the best.
1251504 tur Böylesi daha iyi.
325876 eng That's better.
350332 ita Meglio.
10260361 eng Better.
11165119 sat ᱴᱷᱤᱠ ᱾
1391646 eng Okay.
4726230 spa ¡Órale!
3384686 eng Whoa!
1381630 epo Ho!
404788 por Olá!
924636 yue 你好嗎？
1196342 por Tudo bem?
386479 eng Is that okay?
4772253 jpn いい？
13111715 cmn 可以嗎？
1623504 lzh 可乎？
1623505 eng Is it possible?
1623506 spa ¿Es posible?
4443260 jpn 都合つく？
1012597 fin Ehditkö?
923350 swe Hinner du?
841868 eng Can you make it?
1209377 tur Onu yapabilir misin?
17570 eng Can you do that?
I think I've found myself a new toy — I could happily spend all day finding bizarre little chains between two sentences that seem to have nothing to do with each other.
This is fascinating. Thank you for sharing it! The "indirect links" on Tatoeba have also often reminded me of a huge game of telephone, haha!
A few things stick out to me about the example you posted with "Rome wasn't built in a day". It seems to me that in general, the idioms/proverbs do a lot of the "heavy lifting" in terms of semantic drift. This makes perfect sense in my mind, since it is common on Tatoeba to see translations of idioms/proverbs that have similar figurative meanings but completely different literal meanings. (It could also be that some of these translations are incorrect or not a perfect fit - I know I've definitely made some imperfect translations of idiom/proverbs here.) Seems like proverbs provide a kind of "bridge" for a sentence's meaning to change drastically in just a few links.
It seems like the same applies to very short sentences or exclamations e.g. "Okay." and "Whoa!", since these can be applied in such diverse real-world situations.
This all makes me wonder if some sort of computational technique could be used for "proverb/idiom detection" on Tatoeba sentences. Perhaps these types of sentences could be detected automatically by comparing how well the words in their different translations "map onto each other" one-to-one. That is, if a sentence has 2 distinct translations with completely different/unrelated keywords, then it is more likely to be an idiom or proverb, and this likelihood could maybe be aggregated among all of its translations.
Great insights!
When I looked at the long chains, I did notice those tiny one-word sentences popping up everywhere — “Whoa!”, “Anda.” (Spanish), — but the bit about proverbs being the big semantic drift engines hadn’t crossed my mind. That’s brilliant. You’re right — a proverb will happily hop into another language as a different proverb with the same vibe but no literal overlap, and from there it can wander off in surprising directions.
I’ve also been thinking about how to single out those “gluons” (glue sentences) that disperse meaning the most. Didn’t come up with a perfect recipe, but I did cook up something fun.
I called them Link Gremlins — you’re supposed to roll your “r” in gremlins here for extra menace. I pronounce “Link Gremlins” with a Scottish accent in my head — somehow it makes them sound even more mischievous.
They’re the little rascals sitting in Cloud 1 who link to a ton of other sentences, in a ton of languages, and act like semantic super-spreaders. Remove one and you’d break a lot of the indirect paths in the cloud.
Here’s my current rogues’ gallery of the top 20 gremlins in Cloud 1:
7872529 gos deg=181 langs=56 bridge=0.95 reach2=1425 wc=1 Moi.
373345 fra deg=135 langs=61 bridge=1.00 reach2=1286 wc=2 Bonjour !
373330 eng deg=311 langs=168 bridge=0.99 reach2=1195 wc=1 Hello!
582738 nld deg=72 langs=8 bridge=0.96 reach2=1148 wc=2 Rot op!
2115099 fin deg=65 langs=7 bridge=0.78 reach2=1131 wc=2 Älä viitsi!
373320 eng deg=414 langs=169 bridge=1.00 reach2=1120 wc=3 How are you?
723987 ukr deg=100 langs=28 bridge=0.99 reach2=1111 wc=2 Як справи?
699070 ell deg=53 langs=23 bridge=1.00 reach2=1098 wc=2 Τί κάνεις;
8267435 fra deg=78 langs=5 bridge=1.00 reach2=1077 wc=3 En route !
1101574 nld deg=65 langs=11 bridge=1.00 reach2=1073 wc=2 Ga weg!
373341 epo deg=108 langs=53 bridge=1.00 reach2=1052 wc=1 Saluton!
392055 kor deg=80 langs=43 bridge=1.00 reach2=1036 wc=1 안녕하세요.
1858850 eng deg=121 langs=75 bridge=0.99 reach2=1031 wc=1 Hello.
509819 fra deg=88 langs=44 bridge=1.00 reach2=1021 wc=2 Salut !
4857568 cmn deg=59 langs=41 bridge=1.00 reach2=997 wc=1 你好。
373339 nld deg=60 langs=33 bridge=1.00 reach2=992 wc=1 Hallo.
9025685 fra deg=52 langs=3 bridge=1.00 reach2=990 wc=2 Décampe !
411945 ita deg=73 langs=39 bridge=1.00 reach2=982 wc=1 Buongiorno!
607364 ita deg=71 langs=36 bridge=1.00 reach2=976 wc=1 Ciao!
240634 eng deg=135 langs=34 bridge=0.97 reach2=969 wc=2 Get away!
What the columns mean:
deg → How many direct translations it has — the gremlin’s “friend count”.
langs → How many different languages those direct translations cover.
These two are pretty easy to calculate, and I know CK and sharptoothed post lists of these “highly linked” sentences quite regularly. I just wanted to take it a step further with my own “cloud factor”
bridge → How much of a “bridge” it is between otherwise separate parts of the cloud (1.00 means it’s an excellent connector).
reach2 → How many sentences you can get to in two translation hops
wc → Word count. Most gremlins are short
Basically, these are the little teleporters that let your sentence hop from one corner of the cloud to another in record time, often through phrases so short and vague they can mean a dozen different things depending on the situation.
This analysis isn’t about measuring how far the meaning drifts in just a few hops. I did hear your idea about looking for sentences whose translations use very different keywords — that’s a neat angle, and maybe I’ll poke at it later. That was a nice one!
EDIT – some late thoughts.
After staring at the numbers a bit more, I realised that Link Gremlins are, in the end, mostly the highly linked and highly translated sentences that happen to reach very far in one or two hops. They’re great for connectivity… but not necessarily the best meaning drifters.
For true “meaning drifters” I’ve been toying with another idea. I could come up with a list of, say, a thousand pairs of sentences that, from a purely subjective point of view, seem so wildly different you can’t immediately imagine how they could be linked.
Then I’d:
Run the linking algorithm for each pair.
Gather all the intermediate sentences that appear in these chains.
See which ones turn up most often.
The result would be a kind of “meaning disperser” ranking – not a pure computational metric like the gremlins, but a blend of subjective selection and algorithmic analysis
> For true “meaning drifters” I’ve been toying with another idea. I could come up with a list of, say, a thousand pairs of sentences that, from a purely subjective point of view, seem so wildly different you can’t immediately imagine how they could be linked.
I started it, and it was a pain.
I’d probably grow a beard down to my knees before hitting a thousand.
So, I cheated. I gave ChatGPT a CSV file with all the English sentences from Cloud 1 — about 50,000 of them — and told it to:
pick 1,000 random pairs,
keep only sentences with more than four words, and
use its own mysterious inner “surprisingness” detector to choose them.
It came back with this list:
https://docs.google.com/spreads...it?usp=sharing
I skimmed through it; the pairs do look pretty surprising by my standards too.
Then I ran my translation-path-finder script for each of the 1,000 pairs, saving every sentence that appeared anywhere along those paths.
Here are 13 sentences that appeared in 100 or more chains:
335152|ita|Sono d'accordo.
4726230|spa|¡Órale!
24697|eng|Come on!
268108|eng|All right.
2175|eng|No way!
433859|eng|I've had enough.
373330|eng|Hello!
52022|eng|Terrific!
1018519|fin|Lopeta jo!
241077|eng|Let's go!
695416|tlh|nuqneH?
1381662|epo|Ba!
2115099|fin|Älä viitsi!
Not entirely sure what to make of the list yet, though.
"You and I are very good friends."
->
"She is unconscious."
I’m not saying our friendship is intense, but apparently it’s literally knocking people out.
17603 eng You and I are very good friends.
549442 fra Nous sommes de bons amis.
249493 eng We are good friends.
6779176 nld Wij zijn boezemvriendinnen.
3816927 dsb Smy pśijaśelki.
2107346 eng We're friends.
2458877 eng He and I are friends.
453992 nld Hij is mijn vriend.
2923522 tat Ул минем дустым.
632489 deu Sie ist meine Freundin.
312168 eng She is my girlfriend.
1280528 ita È la mia ragazza.
1804030 eng That's my girl.
1765179 tur O benim kızım.
724401 eng She's my daughter.
1369583 fra C'est ma fille.
7453168 eng That's my daughter.
1377536 por Esta é a minha filha.
4500104 eng This is my child.
1900399 por Este é o meu filho.
1785589 deu Das ist mein Sohn.
2517404 rus Это мой сын.
2339329 spa Es mi hijo.
2002165 tur O benim oğlum.
1804029 eng That's my boy.
1072525 por Esse é o meu garoto!
1020765 spa ¡Ése es mi chico!
51144 eng That a boy!
3488188 jpn よくやった！
4584516 nld Goed gedaan!
395934 eng Great!
4726230 spa ¡Órale!
8267435 fra En route !
1324269 deu Lasst uns gehen.
5824883 fra Allons-y.
1291660 spa Anda.
316925 eng She walks.
11466264 ind Dia berjalan kaki.
8475590 eng He walked.
3808960 ita Ha camminato.
2928468 fra Il a marché.
2928469 ber Yedda.
3259071 eng He left.
1689192 ita È partito.
8666917 rus Он в отключке.
1452761 eng He's out cold.
388191 fin Hän on tajuton.
28206 eng She is unconscious.
Sorry, I’ll stop flooding now — it’s just that this whole “linguistic Chinese Whispers” thing excites me far more than it probably should. And this is the only place I know where there’s even a chance someone else might share that feeling.
I’ll make myself scarce for a bit.
This is really interesting, thank you for sharing!!
El contenido de este mensaje va en contra de nuestras normas y por lo tanto se ha ocultado. Se muestra sólo a los administradores y al autor del mensaje.
#7675229 számú mondat
Eszperantó nap – 2025. július 26.
https://esperantohea.hu/eszpera...025-julius-26/
La 26-a de julio estas la Tago de la Unua Libro pro la aperigo de la unua lernolibro de la Lingvo Internacia + (la Unua Libro) verkita de D-ro Esperanto (kaŝnomo de L. L. Zamenhof). La dato (la 26-an de julio 1887 laŭ Gregoria kalendaro aŭ la 14-an de julio 1887 laŭ la rusa Julia kalendaro) estas tiu de la dua aprobo fare de la rusa cenzuro por la publikigo de la lernolibro, kun la permeso de disvastigo.
+ Международный языкъ. Предисловіе и полный учебникъ.
кто может добавить ягнобский язык и памирские языки?
Ki tudja hozzáadni a jagnobi (1) nyelvet és a pamíri (2) nyelveket?
1.) Tádzsikisztánban élő, 25 ezer lélekszámú nép.
2.) ?
Kiu povas aldoni la jagnoban lingvon kaj la pamirajn lingvojn?
✹✹ Stats & Graphs ✹✹
Tatoeba Stats, Graphs & Charts have been updated:
https://tatoeba.j-langtools.com/allstats/
El contenido de este mensaje va en contra de nuestras normas y por lo tanto se ha ocultado. Se muestra sólo a los administradores y al autor del mensaje.
El contenido de este mensaje va en contra de nuestras normas y por lo tanto se ha ocultado. Se muestra sólo a los administradores y al autor del mensaje.
Hello, Tatoeba was updated today. What’s new?
- Content report: there is a new "flag" button on the top-right of Wall posts and sentence comments to ease the report of inappropriate content to admins. This is a counter-measure to the increased spam Tatoeba is seeing recently, and also a feature some people have been asking for in the past.
- Two new languages have been added: Svan and Ao Naga. Cheers to abiniz and monsen_sanang_ai, respectively, for requesting them! Tatoeba now supports 429 languages.
Learn more about Svan: https://en.wikipedia.org/wiki/Svan_language
Learn more about Ao Naga: https://en.wikipedia.org/wiki/Ao_language
- A total of 50 language icons (flags) should now look nicer. The icons have been updated to a vector-based image instead of a raster image, which means they won’t look pixelated or blurry but always sharp, even when zooming in.
Thanks for the update, Gillux.
If I might use this opportunity to make a suggestion, I think it would be good to implement something that will make it more difficult for spam accounts to register themselves in the first place (captcha, email links etc.?) and delete tens of thousands of existing ones (there are quite possibly over a hundred thousand by now). Nine out of ten newly registered accounts are spammers. Of course, most of them remain silent and passive with ads and links in their profiles. This cannot be not a problem.
I agree. I have been watching the spam accounts come in literally by the minute.
I’ve been reporting spam comments as I see them, but it’s troubling that they can still have a bunch of links in their profile without making any posts. I’d like to help out with that if there’s any way I can. Probably I am too new of a user to have admin privileges, but if it would be helpful, I’m happy to throw together a script that will make a list of likely spam accounts based on common indicators (long link lists in the bio, certain keywords, lack of languages or sentences, etc). Then some admin can run through the list and delete the ones they see fit.
EDIT: just realized that user bios aren’t included in the downloadable data dumps, so maybe that won’t work.
frpzzd, thank you for your help. I have tried to figure out a common feature of spam accounts, but there is nothing really standing out, so I also think we’d need some human verification to confirm account purge.
However, I think it is too early to start deleting spam accounts. It would be better to find a way to stop the influx first, and only then start to cleanup.
As you realized, we do not export user bios. If you are willing to help, I can provide you with a database export of user bios and other information.
Certainly, carrying out any kind of account purge without double-checking each account would be a bad idea. Maybe the job can be streamlined with some analysis of the spam accounts though.
If you think it is too early to begin removing the spam accounts, I guess there's no rush. But in any case, I'd definitely love to take a look at a data dump of user bios if you're willing to share one. Maybe there are some useful features that can be extracted.
Waldelfe, I understand your concern. I am aware of the constant influx of spammy accounts, and be sure that I am willing to stop it. But this is not an easy task.
About adding a confirmation email step, I think it’s is unlikely to help. According to my own research, these accounts are created by humans. They are already going through as many open-registration websites as possible, to create user pages containing a lot of links (see, for example, user xx88art56). Tatoeba is likely the only website in 2025 not requiring email confirmation. If they can do it on other websites which require email confirmation, they will keep doing it on Tatoeba regardless of email confirmation.
About adding a captcha, I don’t think the benefits of the captcha (less spam accounts that are barely noticeable) outweigh the drawbacks (less legitimate users are able to register). Besides, captchas are unlikely to stop the spammers because they are likely real humans.
You are welcome to join the ongoing discussion about prevention of spam accounts on GitHub: https://github.com/Tatoeba/tatoeba2/issues/1613
Why would less legitimate users be able to register if a captcha was added? And thank you for the update!
You are welcome, araneo.
Captcha notoriously lower accessibility: https://en.wikipedia.org/wiki/C...#Accessibility
Captcha can be challenging to solve, even when not visually impaired. If you have poor mental of physical health, simply browsing a website may require a lot of energy; having to solve captchas on the top of that is not helping.
In general, I don’t think anybody is happy to be prompted to solve a captcha.
For these reasons, I’d rather not add more captchas on the web unless there is a really good reason to do so.
Thank you for the explanation and for taking that into consideration!
in my experience captchas are easily solved by modern language models, perhaps even more so than by humans https://www.youtube.com/watch?v=satnl1KTEXM
and even before that there are services employing humans from third-world countries to solve captchas for pennies
one approach i personally like is to hide submissions from new users until at least one sentence/post/whatever from them is approved by a moderator, but that might take effort to implement here since i don't think there's even a way to hide sentences yet
Interesting idea. Alternatively, for an approach that avoids making more work for the mods, we could consider hiding profiles (or just hiding profile bios) of users who have not written any sentences yet by default.
Many spammers seem to not write any sentences at all, and just fill up their bios with links. The spam accounts that write sentences are at least a little more likely to get noticed, flagged and deleted.
> - Content report: there is a new "flag" button on the top-right of Wall posts and sentence comments to ease the report of inappropriate content to admins.
Thanks! That's very useful!
El contenido de este mensaje va en contra de nuestras normas y por lo tanto se ha ocultado. Se muestra sólo a los administradores y al autor del mensaje.