Cyclic linguistic attractors?
« previous post | next post »
A student who's been working on LLM-style AI transformations of symbolically-represented music recently tried mapping a (fragment of a piece) back and forth between two genres, e.g. baroque and pop. She found that after a couple of steps, the results reach a fixed point and don't change any more.
This is different from what used to happen with machine translation between different languages. The Wikipedia article on Round Trip Translation tells us that
Although the use of RTT for assessing MT system quality or the suitability of a text for MT is in doubt, it is a way to have fun with machine translation. The text produced from an RTT can be comically bad. At one time websites existed for the sole purpose of performing RTT for fun. Other variations send the text through several languages before translating it back into the original or continue translating the text back and forth until it reaches equilibrium (i.e., the result of the back translation is identical to the text used for the forward translation). RTT as entertainment appeared in Philip K. Dick's novel Galactic Pot-Healer. The main character runs book titles and sayings through RTT then has his friends try to guess the original. The Australian television show Spicks and Specks had a contest called "Turning Japanese" which used RTT on song lyrics. Contestants needed to correctly guess the title of the song from which the lyrics were taken.
But I wondered whether things might be different now that MT is pretty good, at least between well-document language pairs. So I took the first two sentences from a CNN news story, and ran them back and forth with French via Google Translate — and indeed, after a couple of cycles, the process seems to have converged on a fixed point:
President Donald Trump has fired Pam Bondi as attorney general.
She will be replaced for now by Deputy Attorney General Todd Blanche, who will serve as acting attorney general, Trump said.
Le président Donald Trump a limogé Pam Bondi de son poste de procureure générale.
Elle sera remplacée, pour l'instant, par le procureur général adjoint Todd Blanche, qui assurera l'intérim, a déclaré M. Trump.
President Donald Trump has dismissed Pam Bondi from her position as Attorney General.
She will be replaced, for the time being, by Deputy Attorney General Todd Blanche, who will serve as interim, Mr. Trump stated.
Le président Donald Trump a relevé Pam Bondi de ses fonctions de procureure générale.
Elle sera remplacée, pour le moment, par le procureur général adjoint Todd Blanche, qui assurera l'intérim, a déclaré M. Trump.
President Donald Trump has relieved Pam Bondi of her duties as Attorney General.
She will be replaced, for the time being, by Deputy Attorney General Todd Blanche, who will serve as interim Attorney General, Mr. Trump stated.
Le président Donald Trump a relevé Pam Bondi de ses fonctions de procureure générale.
Elle sera remplacée, pour le moment, par le procureur général adjoint Todd Blanche, qui assurera l'intérim, a déclaré M. Trump.
President Donald Trump has relieved Pam Bondi of her duties as Attorney General.
She will be replaced, for the time being, by Deputy Attorney General Todd Blanche, who will serve as interim Attorney General, Mr. Trump stated.
I don't have time this afternoon to check whether the same thing happens with longer cycles, other passages, longer passages, other languages, other translation systems, three- or four-way language cycles, etc. — readers are welcome to try.
But I suspect that many similar experiments are now similarly likely to converge on cyclic linguistic attractors in machine-translation space…
Benjamin Geer said,
April 3, 2026 @ 6:12 am
News stories probably aren’t the ideal input for this, because for a long time now they’ve been written in a kind of “translationese” which is very formulaic and in which every word or phrase has a standard equivalent in many languages. This is particularly true of wire service reports, which are published in multiple languages and which LLMs have certainly been trained on.
languagehat said,
April 3, 2026 @ 8:54 am
For "different from" read "much like"?
MattF said,
April 3, 2026 @ 9:38 am
I wonder, though, whether the translation after convergence is superior to the original translation. It’s well understood, for example, that many numerical summations can be rearranged to converge to a variety of different values.
Allan from Iowa said,
April 3, 2026 @ 9:59 am
I had always thought that a rapid convergence was a sign that the translations were high quality.
But after reading Benjamin Geer's comment, it's clear that just because round-trip translations of a news story converge, that doesn't mean that the same translation software would yield similar results for a more literary or more colloquial text.
Allan from Iowa said,
April 3, 2026 @ 10:08 am
Sometimes a round-trip translation is an improvement.
Once I saw a text on the web that apparently had been auto-translated from Turkish to English. The English version was full of apparently misspelled Turkish words. I used Google Translate to convert it back to Turkish, which changed some of the spellings (mainly dotted versus undotted I), and then back to English again, yielding a more understandable result.
Jerry Packard said,
April 4, 2026 @ 9:42 am
Another interesting phenomenon is pitting 2 ai platforms against each other. The result is an endless ‘loop’ (not a recursive loop really) of never-ending discussions unless you embed somewhere in the algorithm(s) a ‘goal’ or place to stop.