“I have gone into my own way”

« previous post | next post »

In a series of recent posts we’ve explored the fun side of recursive weighted sums and point nonlinearities as a translation algorithm: “What a tangled web they weave“, “A long short-term memory of Gertrude Stein“, “Electric sheep“, “The sphere of the sphere is the sphere of the sphere“. But the featured translations have all involved inputs of characters in kana, hangul, Thai script, and other non-Latin alphabets, and it’s natural to wonder whether this is an essential part of the game.

No — here are various repetitions of “è “, “îî “, and “îè ”  translated from Greek:

è è è è è è è è è è Things to Do
è è è è è è è è è è è Date of Issue No.
è è è è è è è è è è è è May 2009
îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî I have forsaken myself for it to be with you
îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî îî I have resuscitated myself for my own sake I have forgiven myself for myself
îè îè îè îè îè îè îè îè îè You’re going to be yours
îè îè îè îè îè îè îè îè îè îè îè îè îè You’ll be out of your way
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You’re on your way out of the sun
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You’re on your way back to your day
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You are on your way back to the day you are in your country
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You have been signed in. You have signed in. You have signed in.
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You are on your way to the last day of your stay. You have reached the last day of your stay.
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You have finished your call and have signed in.
îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè îè You have been signed in. You have made a call. You are on your way. You are on your way. You have signed in.

Since it’s somewhat unfair to ask the program to translate from a language that lacks the characters we input, let’s let Google detect the language. Here’s “Romanian – detected”:

î î î î î I’m going
î î î î î î î î î I’m in
î î î î î î î î î î I’m walking around
î î î î î î î î î î î I’m in my way
î î î î î î î î î î î î î î î î î î î î î I’m in my heart
î î î î î î î î î î î î î î î î î î î î î î î I’m starting to get in my heart
î î î î î î î î î î î î î î î î î î î î î î î î I’m in my own time
î î î î î î î î î î î î î î î î î î î î î î î î î I’m starting to get around in my life
î î î î î î î î î î î î î î î î î î î î î î î î î î î î î î î I have gone into my own way

The same:

îü îü îü îü îü It will be yours
îü îü îü îü îü îü îü îü I will have it
îü îü îü îü îü îü îü îü îü I will be in my heart
îü îü îü îü îü îü îü îü îü îü îü îü îü I will have it for myself
îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü I will have her
îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü I will have my bloom
îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü îü I will have my blessing I will have my blessing

And here’s “Hawaiian – detected”:

ō͂ Larger
ō͂ō͂ northshore
ō͂ō͂ō͂ō͂ Espanyol
ō͂ō͂ō͂ō͂ō͂ō͂ edit
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Yeah
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Armin
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ onestic
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Other
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Wedding
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ ich
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ aircraft
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Best
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ wipers
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ edging
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ Love
ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ō͂ around

I haven’t been able to create similar effects with pure ASCII gibberish as input — perhaps someone else can figure out how to do so.



20 Comments

  1. Johan P said,

    April 27, 2017 @ 10:01 am

    You mean, without the diacritics?

    “sa sa sa sa sa sa sa sa sa sa sa sa sa sa” becomes (in “romanian”) “Be sure to get it to him”. “mo mo mo mo mo mo mo mo mo mo mo” becomes (in “japanese”) “It is also possible”. “ty ty ty ty ty ty ty ty ty ty” (in “vietnamese”) becomes “You are the Company”. “bo bo bo bo bo bo bo bo” (in “Swedish”) becomes “Top above top of the top”.

    [(myl) Nice!]

    Or do you mean entirely random? That seems harder.

    [(myl) Yes, keyboard banging generally produces something interesting from a Thai keyboard, but not from an American English one…]

  2. Robert Ayers said,

    April 27, 2017 @ 10:14 am

    Doesn’t require odd words, doesn’t require diacritics, doesn’require unusual languages. I discussed sequences of “he” — first short plain word I tried — “translated” into Spanish back on 18 April and commented in the “Electric Sheep” article.

  3. Peter Erwin said,

    April 27, 2017 @ 10:32 am

    Someone in one of the earlier posts mentioned that “plume” repeated several times would generate odd results. Other repetitions of short ASCII strings can produce interesting bits of weirdness as well.

    For example, “cat” repeated several times (detected as “Vietnamese”) just produces copies of the input… until you get up to seven or more repetitions, where it starts inserting “1” and “2” between instances of “cat”. After about 15 repetitions, you start getting random bits of HTML and/or email addresses. For example, 16 repetitions translates as “Catcat_cat_html-loader_html”, and 25 repetitions is “Catcatcatcat@gtiksdh.net”.

    Repetitions of “bing” (detected as “Chinese”) get translated as one or more instances of “ice”; “bing” repeated 13 times is “Athenian ice and ice”, 18 times = “The other parts of the world”, and 31 times is “An endless stream of ice and ice bingbingbingbingbingbing”, which seems oddly apropos. (And then 32 times = “All the people who have made their brains”)

    The shortest instance I discovered is “armarm”, which is apparently Latin for “The refrigerator”. (“arm” repeated 10 or 12 times, on the other hand, is apparently Hindi for “armor monarchy”)

  4. Ethan said,

    April 27, 2017 @ 10:39 am

    Johan P: I cannot comment about what happens when the input language is set to Romanian or Swedish, but for Japanese the ascii input characters are treated as a romaji representation of Japanese kana. So I think this simply adds a pre-processing step to the original tests with repeating kana. By the way, “momomomomo…” is arguably possible in Japanese, leading for instance to a manga series Sumomomo Momomo: The Strongest Bride on Earth.

  5. Sergey said,

    April 27, 2017 @ 10:46 am

    For what I remember from a lecture on machine translation, it works approximately like this: for each word, itself and the set of surrounding words, and maybe also of the previous words in the translation is fed to the neural network that produces the most probable translated word. That the first words keep changing even when a word is added to an already long sequence suggests that they use a quite wide context. Or maybe they use some kind of postprocessing layer that makes sure that the grammar in the result is put together right (looks very likely for the languages with declensions but not sure if they do something of the sort for English too), or maybe even some kind of two-pass processing when the translation from the first pass is used as a part of word context for the second pass. I wonder if some experiments can be designed to deduce the logic.

    These neural networks are huge: each word has its own probability input and output (although for the input the value is more of a binary: the word is either present or not). So for 10K words that would be a network with 10K outputs, and 10K inputs for the current word, then 10K inputs for the previous word, 10K input for the next word, and if the translated context is used, 10K inputs for the previous translated word, so the context extension gets pretty expensive pretty fast.

    That the random sequences of characters produce some output at all, and a varying one to boot, suggests that there probably is a layer that tries to correct the mistakes in the input, similarly to the suggestions when typing in a search request. This layer might be going not word-by-word but character-by-character. And then the result of this layer would be fed to the translation layer, so it should produce even more variability.

  6. Gwen Katz said,

    April 27, 2017 @ 10:49 am

    I think it made a rather nice poem.

  7. Johan P said,

    April 27, 2017 @ 10:54 am

    @Ethan Sure, I get that. I should really have excluded that example, which was autodetected. As a Swedish speaker I can tell you that “bo” is either a verb meaning “reside”, a noun meaning “nest” or a given name, and has no relation to the suggested sentence, though.

  8. Alon said,

    April 27, 2017 @ 11:06 am

    Works with il and un.

    [(myl) Nice! The obligatory screenshots:

    It’s especially interesting that “un un un un un un un un un un un un un un un un un un un un un un” translates as “One of the largest united states of the world.” ]

  9. J. Goard said,

    April 28, 2017 @ 12:42 am

    Values of n for which Romanian>English does not translate n repetitions of “da” as “n” repetitions of “yes”, but instead leaves only the second one as “da”: 11,13,14,15,16,19,21,23,31,33,37,43,45+

    Here’s the beginning of the German>English:
    da > there
    da da > since
    da da da > da da da
    da da da da > Since there da da
    da da da da da > Since there da da da
    da da da da da da > Since there da da da da
    da da da da da da da > Because there da da da da
    da da da da da da da da > Because there da da da da da
    da da da da da da da da da > Since there da da da da da da

    Italian>English is “from” for n=1, original string of “da”s only for n=7, and string of “yes”s for all other n. However, when just the “a” in the last “da” is deleted, it changes radically to stuff like this:
    “From on by by by by by by by by by by by by by by by by the way of the da”
    (But only for some n. )

  10. John Swindle said,

    April 28, 2017 @ 1:23 am

    Romanian “îè îè îè îè îè îè îè îè îè îè îè îè îè” for “You’ll be out of your way” and Greek “î î î î î î î î î î î î î î î î î î î î î î î î î î î î î î î” for “I have gone into my own way” appear to be semantically related. If we looked further we’d probably find that in Proto-Balkan “every one to his own way” was expressed as something intermediate between the two.

  11. J. Goard said,

    April 28, 2017 @ 7:23 am

    Some interesting English > Korean:

    do > 해야 할 것
    [thing that must be done]
    do do > 그러다.
    [then; if; in that way]
    do do do > 해라.
    [do it! — NO POLITENESS SUFFIX]
    do do do do > 어떻게해야합니까?
    [How must it be done? — FORMAL SUFFIX]
    do do do do do > 어떻게해야합니까?
    do do do do do do > do do do do do
    [Yes, one “do” just gets deleted.]
    do do do do do do do > 해야 할 일
    [work that must be done]
    do do do do do do do do > do do do do해야 마
    [weird garbling of grammar for ‘must’ and ‘don’t’, I guess]
    do do do do do do do do do > do do do do do do해야 마
    do do do do do do do do do do > 해야합니까합니까합니까합니까합니까합니까합니까
    [must I do it do it do it do it, etc.]

  12. Johan P said,

    April 28, 2017 @ 8:32 am

    Question: Has anyone been able to produce this with “syllables” of three letters or more? Or with alternating syllables? I can’t get any to work.

  13. guilty bystander said,

    April 28, 2017 @ 8:36 am

    I tried it with Finnish “mä”:
    Me
    Me
    I am here
    I am here
    I am here
    I am here
    I do not care
    I do not care
    I do this any more
    I do not care
    I do not care
    I am going to do so
    I will do so as soon as possible
    I am going to do it all the time
    To do more than once
    If i do not hesitate
    To do more and more
    To do more than once
    More and more of this
    More if you do not know me
    More if you do not know me
    More and More to be Here
    More and More to be Here
    More if you do not know what to do
    More if you do not know what to do
    More if you do not know what to do

  14. Johan P said,

    April 28, 2017 @ 8:38 am

    After trying a few dozen I got one, in “japanese” which doesn’t really count, I suppose.

    dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa dasa

    —>

    Sorry, it’s still in good condition and it’s about to be solved.

  15. Vulcan With a Mullet said,

    April 28, 2017 @ 9:14 am

    I just realized – this isn’t word salad generated by an algorithm, it’s a sign that artificial intelligence is finally self-aware and is attempting to express itself through surrealist poetry!
    Better than taking over the world and enslaving humanity.. I SUPPOSE….

  16. Yet Another John said,

    April 28, 2017 @ 8:25 pm

    Entering many space-separated repetitions of “que” and “translating” from Spanish to English is interesting. (To be clear, I mean “que” and not “qué.”)

    Repeat it ten times, and you get the usual weirdness, like “I would rather have than that they would have to find that which”

    But if you repeat it 60 or more times, the English translation will begin with text apparently copied from a TripAdvisor review:

    “I do not recommend this hotel for Young singles, Great people, Great pool Was this review helpful? Yes View profile Send message …”

    Curiouser and curiouser.

  17. Rebecca said,

    April 28, 2017 @ 9:26 pm

    Detected as Polish:
    wy
    wywy
    wywywy
    wywywywy
    wywywywywy
    wywywywywywy
    wywywywywywywy
    wywywywywywywywy
    wywywywywywywywywy
    wywywywywywywywywywy
    wywywywywywywywywywywy
    wywywywywywywywywywywywy

    Translated to English:
    you
    Wywy
    Outgoing
    Off road
    Expedition
    Expulsive
    Food
    Food
    Food and drink
    Food and drink
    Food and drink
    Food and drink

  18. Keith Ivey said,

    April 29, 2017 @ 9:09 am

    If you translate “thethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethethe” from English to German, you get “Wörterbuch Wörterbuch Wörterbuch” (as I mentioned on Language Hat).

  19. Smut Clyde said,

    April 30, 2017 @ 1:52 am

    “syllables” of three letters or more?

    Taking ‘tää’ (from the Ur Sonata) and repeating it increasingly many times provided a kind of Beckett / Ionescu collaboration:
    Come on
    Here is the place
    Here’s what to do
    Be sure to do this
    Keep this in mind
    Keep this in mind
    Be sure to keep this in mind
    Be sure to keep this in mind
    Keep this in mind
    Keep this in mind
    Must take into account the need to
    Must take into account the need to take account of
    Taking account of the need to take account of
    Taking into account the need to take account of
    Taking into account the need to take account of
    It is also necessary to take account of the fact that
    Taking into account the need to take account of
    Taking into account the need to ensure that there is no

  20. Smut Clyde said,

    April 30, 2017 @ 1:59 am

    The results of iterated Lovecaftian invocations are recognised as Finnish, and are just as scary as you might expect. What do the Finns know that the rest of us don’t?
    http://eusa-riddled.blogspot.co.nz/2017/04/invocation.html

    Yeah! Shub-Niggurath!
    Come on! Shub-Niggurath!
    Come on! Shub-Niggurath!
    YEARS! Shub-Niggurath!
    Shub-Niggurath!
    We have Shub-Niggurath!
    THANK YOU NOW! Shub-Niggurath!
    We have Shub-Niggurath!
    THANK YOU NOW! Shub-Niggurath!
    Shub-Niggurath!
    Shub-Niggurath!
    Shub-Niggurath, ever since!
    Shub-Niggurath! All right now!
    Shub-Niggurath! Everyday!
    Shub-Niggurath ever had it!
    Shub-Niggurath! Ever since!
    Shub-Niggurath has ever been here!
    Shub-Niggurath has always been here!
    Shub-Niggurath has ever been here!
    Shub-Niggurath has ever been here in all times!
    Shub-Niggurath! Ever since, all around the world!
    Shub-Niggurath! In the days of nowadays, all over the world!
    Shub-Niggurath! EVERYTHING EVERYTHING ALWAYS ALLOWED!
    Shub-Niggurath! Ever since, all over the world!
    Shub-Niggurath! Ever since, all around the world!
    EVERYTHING EVERYTHING IS HERE! Shub-Niggurath!
    Shub-Niggurath! Ever since, all around the world!
    Shub-Niggurath! EVERYTHING EVERYTHING ALWAYS ALLOWED!
    Shub-Niggurath! Ever since, all around the USA!
    Shub-Niggurath! EVERYTHING EVERYTHING ALWAYS HERE!
    Shub-Niggurath! EVERYTHING EVERYTHING ALWAYS ALLOWED NOW!
    Shub-Niggurath! Ever since, all around the world!
    Shub-Niggurath! EVERYTHING EVERYTHING!
    Shub-Niggurath! Ever since, all around the world!

RSS feed for comments on this post