Ages and ages ago a group of colleagues of mine did a study of faxing Arabic-language prescriptions- the low resolution of the fax machine, as well as the textured appearance of the forgery-resistant paper that prescriptions had to be printed on, meant that diacritic marks got smudged in ways that were leading to clinically-meaningful confusion on the part of the reader. Here's their paper about it: https://pubmed.ncbi.nlm.nih.gov/22178294/
I don't speak Farsi, but some Arabic, and the Farsi alphabet is related.
The letters are completely different in pronunciation: with one dot, it's like the English letter B, and with two it's either like Y or the pronunciation of double E.
The pronunciation equivalent is called a 'minimal pair' in linguistics. Two words that differ by a single sound, and they're pretty common.
The written equivalent is probably more common in most other languages than in English. English has a relatively 'deep' orthography compared to most languages, ie. the spelling makes no gosh darn sense.
An English example (when spoken) might be 'scanning' versus 'skinning', with very different implications. A vet might scan a dog, but hopefully they wouldn't skin one. I have no idea what 'skinning electron microscopy' could be, either!
The presence of the single dot determines how to read the grapheme, ie which specific letter of the alphabet it is.
A missing dot over the “i” in the Latin alphabet is unambiguously a mistake (depending on the font or script style). But if you missed the horizontal crossbar within the letter “t”, you would read it as an “l”. For example, confusing “mate” as “male”.
There are, unsurprisingly, many examples of the same thing happening in English in the siblings of my post here.
This is a characteristic of grammatical systems. This contrasts non-grammar-based communication systems, such as the ones used by many animals, where a sound may indicate "a threat is near", and the intensity, duration, or repetition of the sound may encode the proximity/relevance of the threat. This is obviously a very sensible way to operate, especially for neural-net based intelligences, so much so that this still lives inside our language usage as well. Think of the progression of calling someone's name as you try harder to get their attention; same sounds, just louder and longer as you step up the intensity.
However it has problems with the fact that there are only so many sounds that can be modulated like that and still be distinct. Each individual sound consumes not just its "base" representation, but all the possible modulations.
Grammar-based systems pack a lot more meaning into a much tighter encoding, but basically by the pigeon-hole principle, you can't help but end up with radically different meanings next to each other. We still only have so many sounds we can make. In fact such systems kind of want that outcome; per sindriava's example, it is generally good that "fart" and "tart" are fairly close to each other. They will not generally be truly confused for each other, so packing them close together is cheap. Similarly for packing "scanning" and "vegetative" together; no native Farsi speaker would do much more than momentarily blink at such a change, and maybe get a chuckle out of it. Things like "can" versus "can't" cause us much more real-life problems, where they are diametrically opposed to each other but distinguished only by a couple hundred milliseconds of sound between them, and by their grammatical nature, are almost always legal sentences when interchanged. (Maybe it's "always" but you learn to be careful about claiming "always" about any language issue. If anyone can come up with one have fun, though bear in mind you need to be using the correct meanings; "tin can" versus "tin can't" is something else entirely, that "can" is a different word for the purposes of this conversation.)
One of the general lessons that I think comes up in learning other languages is that there aren't that many features that are truly unique to a given language. Ratios vary. Sometimes one language will make heavy use of a feature, like Chinese and tonality, and another will make very, very weak usage of it... but there are English words that are spelled the same and when spoken, are distinguished by tone. By no means is English a "tonal language", tone in English is mostly reserved for emoting, but it's not quite absent. 50 years ago I'd have said ideographs were certainly isolated to certain East Asian languages but now English has some rudimentary ideographs in it, perhaps most notably the eggplant and peach emoji which have widely (albeit not universally) understood meanings now largely severed from their original graphical representations in a very similar manner to ideograms. Much of the time whatever "weird" things you see in a language that it uses widely are still present in your own, just used much, much less.
Given how often I've heard this story over the last few months, I'd half expect the total number of papers to increase with research about this topic itself.
This seems like another example of the 'tortured phrases' [1], where AI-generated or machine-translated gibberish slips into papers. It’s worrying how often these nonsensical phrases bypass peer review scrutiny.
This illustrates to me how amazing LLMs are at 'remembering' _extremely_ obscure information. Somehow LLMs can repeat mistakes made a couple times looooooong ago.
Fascinating that in Farsi "scanning" and "vegetative" differ only by a single dot in the word!
Kinda blows my mind, but then I am only really familiar with English and basic knowledge of a few other European languages.
Anyone here got any more details on this? To me it seems like its as minor as just like missing a dot on an I but totally changes the meaning?
Ages and ages ago a group of colleagues of mine did a study of faxing Arabic-language prescriptions- the low resolution of the fax machine, as well as the textured appearance of the forgery-resistant paper that prescriptions had to be printed on, meant that diacritic marks got smudged in ways that were leading to clinically-meaningful confusion on the part of the reader. Here's their paper about it: https://pubmed.ncbi.nlm.nih.gov/22178294/
Think about the difference between "tart" and "fart".
This is a good example - you just flipped the t upside down.
If you are James Joyce, you are interested in both.
I was thinking more like "Ear" and "Car" or "till" and "tilt"
fuck luck
cant cunt
I don't speak Farsi, but some Arabic, and the Farsi alphabet is related.
The letters are completely different in pronunciation: with one dot, it's like the English letter B, and with two it's either like Y or the pronunciation of double E.
"Let's eat, grandma"
vs
"Let's eat grandma"
"The panda, eats, shoots, and leaves"
vs.
"The panda eats shoots and leaves"
Or the more vulgar
"I helped my uncle, Jack, off a horse"
vs
"I helped my uncle jack off a horse"
Great band!
The pronunciation equivalent is called a 'minimal pair' in linguistics. Two words that differ by a single sound, and they're pretty common.
The written equivalent is probably more common in most other languages than in English. English has a relatively 'deep' orthography compared to most languages, ie. the spelling makes no gosh darn sense.
An English example (when spoken) might be 'scanning' versus 'skinning', with very different implications. A vet might scan a dog, but hopefully they wouldn't skin one. I have no idea what 'skinning electron microscopy' could be, either!
The presence of the single dot determines how to read the grapheme, ie which specific letter of the alphabet it is.
A missing dot over the “i” in the Latin alphabet is unambiguously a mistake (depending on the font or script style). But if you missed the horizontal crossbar within the letter “t”, you would read it as an “l”. For example, confusing “mate” as “male”.
A missing dot over the "i" in the Latin alphabet is not a mistake in Turkish. It's a different letter.
https://haacked.com/archive/2012/07/05/turkish-i-problem-and...
There are, unsurprisingly, many examples of the same thing happening in English in the siblings of my post here.
This is a characteristic of grammatical systems. This contrasts non-grammar-based communication systems, such as the ones used by many animals, where a sound may indicate "a threat is near", and the intensity, duration, or repetition of the sound may encode the proximity/relevance of the threat. This is obviously a very sensible way to operate, especially for neural-net based intelligences, so much so that this still lives inside our language usage as well. Think of the progression of calling someone's name as you try harder to get their attention; same sounds, just louder and longer as you step up the intensity.
However it has problems with the fact that there are only so many sounds that can be modulated like that and still be distinct. Each individual sound consumes not just its "base" representation, but all the possible modulations.
Grammar-based systems pack a lot more meaning into a much tighter encoding, but basically by the pigeon-hole principle, you can't help but end up with radically different meanings next to each other. We still only have so many sounds we can make. In fact such systems kind of want that outcome; per sindriava's example, it is generally good that "fart" and "tart" are fairly close to each other. They will not generally be truly confused for each other, so packing them close together is cheap. Similarly for packing "scanning" and "vegetative" together; no native Farsi speaker would do much more than momentarily blink at such a change, and maybe get a chuckle out of it. Things like "can" versus "can't" cause us much more real-life problems, where they are diametrically opposed to each other but distinguished only by a couple hundred milliseconds of sound between them, and by their grammatical nature, are almost always legal sentences when interchanged. (Maybe it's "always" but you learn to be careful about claiming "always" about any language issue. If anyone can come up with one have fun, though bear in mind you need to be using the correct meanings; "tin can" versus "tin can't" is something else entirely, that "can" is a different word for the purposes of this conversation.)
One of the general lessons that I think comes up in learning other languages is that there aren't that many features that are truly unique to a given language. Ratios vary. Sometimes one language will make heavy use of a feature, like Chinese and tonality, and another will make very, very weak usage of it... but there are English words that are spelled the same and when spoken, are distinguished by tone. By no means is English a "tonal language", tone in English is mostly reserved for emoting, but it's not quite absent. 50 years ago I'd have said ideographs were certainly isolated to certain East Asian languages but now English has some rudimentary ideographs in it, perhaps most notably the eggplant and peach emoji which have widely (albeit not universally) understood meanings now largely severed from their original graphical representations in a very similar manner to ideograms. Much of the time whatever "weird" things you see in a language that it uses widely are still present in your own, just used much, much less.
Some related past threads,
https://news.ycombinator.com/item?id=43759130 ("A weird phrase is plaguing scientific papers (theconversation.com)" — 38 comments)
https://news.ycombinator.com/item?id=35423290 ("Chemist Rafael Luque suspended without pay for thirteen years (elpais.com)" — 120 comments)
https://news.ycombinator.com/item?id=42759862 ("Tomatoes roaming the fields and canaries in the coalmine (deevybee.blogspot.com)" — 33 comments)
https://news.ycombinator.com/item?id=28107614 ("Tortured phrases: A dubious writing style emerging in science (nature.com)" — 259 comments)
https://news.ycombinator.com/item?id=28137980 ("A dubious writing style emerging in science (arxiv.org)" — 40 comments)
>As of today, "vegetative electron microscopy" appears in 22 papers, according to Google Scholar.
Is there any reason that journals shouldn't retract those 22 papers? I suppose those are probably pretty suspect journals in the first place.
So i just chose two at random now and they don't appear to be predatory journals, just journals with Impact Factors lower than 1.
Imagine to retract automaticaly all papers that confuse "then" and "than".
My English pronunciation is so bad that I'd never confuse them, but I see that mistake quite often.
This isn't a mistake though, right? These papers were authored by a chat bot, so why would anyone assume that any of the conclusions are legitimate?
This is the 2017 article with the typo https://jfisheries.ut.ac.ir/article_74390_fa7bb4e3378a7fa45e...
This is the bad translation https://jfisheries.ut.ac.ir/article_74390.html
They probably reviewed the original article in Farsi, and nobody took a deep look at the translated abstract.
This was seemingly unintentional, but I bet there are "trap streets" already in the models that use this principle.
Given how often I've heard this story over the last few months, I'd half expect the total number of papers to increase with research about this topic itself.
This seems like another example of the 'tortured phrases' [1], where AI-generated or machine-translated gibberish slips into papers. It’s worrying how often these nonsensical phrases bypass peer review scrutiny.
[1] https://dbrech.irit.fr/pls/apex/f?p=9999:5::::::
Previously: https://news.ycombinator.com/item?id=43022628
Vibe microscopy is here
Vegetative Electron Microscopy is now the name of my new Psychedelic Darkwave Doom Metal project.
To be known as "The Vegs" for short.
This illustrates to me how amazing LLMs are at 'remembering' _extremely_ obscure information. Somehow LLMs can repeat mistakes made a couple times looooooong ago.
The only thing left to do is pipe the output of a coma patient in an fMRI into the controls of an electron microscope.
https://www.wired.com/2009/09/fmrisalmon/
Shouldn't they start with dead salmon?