ChatGPT writes Haiku

December 21, 2022 @ 9:36 pm · Filed by under Artificial intelligence, Language and literature

« previous post | next post »

[This is a guest post by Bill Benzon]

I’ve been spending a LOT of time with ChatGPT. So naturally, I decided to have it create some haiku.  [VHM:  See the link to Bill's blogpost after the page break.]  This post is about that, but also about a most remarkable woman, Margaret Masterman (1910-1986). She’d studied with Wittgenstein in the 1930s and then went on to create the Cambridge Research Unit in Linguistics in the 1950s. There she became one of the founders of computational linguistics and had a computer generate haiku in 1969. As far as I know, it’s the first time that’s been done.
 
Take at look at the very end. I’ve taken to closing my dialogs by thanking ChatGPT. I know it’s not conscious, nor sentient, but why not? It’s fun. This time I decided to thank it in Japanese. Except that I neither speak nor read Japanese. But I can use Google Translate. I thought ChatGPT would have no trouble, but I do think its reply was rather clever.
 
Best of the season to you, and the rest of the Log.

"Margaret Masterman, pioneering computational linguist [+ Haiku]" (12/17/22)

VHM:  This morning, around the same time that I received Bill Benzon's message with his blogpost about ChatGPT-generated haiku, I received a message from Don Keyser about human generated haiku that brought this article:

"The Capacious Haiku", Janine Beichman, Culture, nippon.com (12/21/22)

Here are examples of each type:

 

ChatGPT

The linguist's mind at work 
Masterman's haiku creation 
A poetic masterpiece

 

Matsuo Bashō (1644–94)

Octopus pots—

brief dreams beneath

the summer moon

—-

Takotsubo ya hakanaki yume o natsu no tsuki

蛸壺やはかなき夢を夏の月    松尾芭蕉

[This] poem is by Matsuo Bashō (1644–94), the greatest haiku poet of all. Octopus pots are ceramic pots that take advantage of the penchant of octopuses for squeezing into small spaces. The fishermen lower the pots to the bottom of the sea during the day and then pull them up shortly before dawn. In the darkness of the pot as it rests on the floor of the sea, the octopus settles in for the night, innocently dreaming, or so the poet imagines. At dawn, when it is hauled up, its brief dream—for the summer nights are short—will be over. Our life, its life, is no more than a dream. But it has been touched by the moon, symbol of enlightenment and eternity.

ChatGPT writes extensively about the nature of haiku and the writing of haiku, Masterman's contributions to the field of machine translation, and so on.  Now I would like to see it describe, explain, and critique its own haiku.  This is something that a human could readily do.

 

Selected readings

December 21, 2022 @ 9:36 pm · Filed by under Artificial intelligence, Language and literature


15 Comments

  1. Laura Morland said,

    December 22, 2022 @ 12:29 am

    Just writing to thank you for introducing me to ChatGPT.

    I've taken up the challenge to write a hymn, and last night I asked ChatGPT to do the same.

    Its second attempt (following more spécific instructions), while not à work of art, was nonetheless astonishing.

    I plan to submit it anonymously to the compétition. (And I thanked ChatGPT, too.)

  2. Victor Mair said,

    December 22, 2022 @ 7:02 am

    @Laura Morland

    After you submit your hymn to the compétition, please share it with us.

  3. Leanne Ogasawara said,

    December 22, 2022 @ 9:48 am

    This was such a fascinating post. Coincidentally I just finished reading Janine Beichman’s marvelous essay and loved your last challenge about being able to reflect and analyze one’s own translation and the original poem… great post!

  4. Jon W said,

    December 22, 2022 @ 10:19 am

    ChatGPT's answer to a request to describe, explain, and critique the haiku is at https://www.dropbox.com/t/W7C3i9Zts2wz13kp (link should be good for the next week). It's not hugely impressive, to my eye, but it's all a matter of what your expectations are.

  5. Taylor, Philip said,

    December 22, 2022 @ 10:30 am

    ChatGPT, in its critique of its own work, writes "This haiku is written in the traditional form of a Japanese haiku, which consists of three lines with a syllable count of 5-7-5". That is what I had expected to find when I looked at its haiku, but no matter how many times I look at, I cannot find that pattern. The pattern that I see is as follows :

    The linguist's mind at work (1+2+1+1+1 = 6)
    Masterman's haiku creation (3+2+3 = 8)
    A poetic masterpiece (1+3+3 = 7)

    What am I missing ?

  6. Jon W said,

    December 22, 2022 @ 10:49 am

    Philip Taylor — ChatGPT can parrot a description of haiku, but its training data have given it only a general idea of what haiku look like. So the poems it generates are close to English-language haiku, even though they're not exact.

  7. Taylor, Philip said,

    December 22, 2022 @ 11:09 am

    It's not so much ChatGPT's haiku that is my concern, more its critique — it claims that the haiku follows the traditional 5-7-5 on/syllable pattern, but I cannot find that pattern in it …

  8. Jon W said,

    December 22, 2022 @ 11:35 am

    ChatGPT doesn't think or reason. It says that haiku follows a 5-7-5 pattern because there are many text strings in its training database stating that haiku follows a 5-7-5 pattern. It has no ability to notice inconsistencies between one sentence in its output and another.

  9. Rob Grayson said,

    December 22, 2022 @ 1:02 pm

    @Taylor, Philip – I gave ChatGPT a topic and asked it to write some Shakespearean blank verse. It came up with something that sort of looked like Shakespearean blank verse, except that some of the couplets rhymed and few of the lines had ten syllables. I told it to remove the rhymes and make each line have ten syllables; it was able to reduce (if not eliminate) the use of rhyming, but the syllable counts were still all over the place. So I asked it to put a syllable count at the end of each line. It proceeded to place a digit at the end of each line, but these digits in no way matched the actual number of syllables in each line.

    I concluded that ChatGPT (i) has no ability to compute or count the number of syllables in a given string of text and (ii) sometimes acts as though it has complied with a reasonably straightforward instruction ("make each line have exactly ten syllables") when it is, in fact, categorically unable to do so. (In the latter case, I would much prefer it to tell me that it's unable to reliably follow the instruction.)

  10. Bill Benzon said,

    December 22, 2022 @ 2:38 pm

    Guys, Right. ChatGPT has difficulty counting. I asked it to count number of words in some sentences and it got the counts wrong, even allowing for the fact that it treats periods as words, something it was kind enough to inform me. I haven't tried arithmetic, but there is a discussion in the literature about large language models being challenged by grade school arithmetic.

    And, since it has no source of "ground" truth, it doesn't know what it doesn't know. It's all text all the way down.

  11. Bill Benzon said,

    December 22, 2022 @ 7:53 pm

    Jon W: I went to ChatGPT and gave it the same prompt you did:

    Describe, explain, and critique the following haiku:

    The linguist's mind at work
    Masterman's haiku creation
    A poetic masterpiece

    Here is its response:

    This haiku is written in the traditional 5-7-5 syllable count and appears to be written in praise of a linguist's mental processes and the art of haiku creation. It mentions "Masterman," who may be a reference to a specific person or a fictional character. The final line, "A poetic masterpiece," suggests that the haiku creation being referred to is highly skilled and well-crafted.

    Overall, the haiku appears to be positive in tone and celebrates the beauty and complexity of language and poetry. However, without more context, it is difficult to fully understand the intended meaning or significance of the haiku. Haiku is a traditional form of Japanese poetry that is characterized by its use of sensory language and specific syllable counts. It often reflects on nature or the passing of time and aims to convey a moment of insight or enlightenment in a brief and concise manner.

    That's rather different from the response it gave you, which doesn't surprise me at all. Is it better or worse? I can't say that I care. What's interesting to me is that it can perform that task at all and that it has some 'discretion' in just how it respondes.

  12. Richard M said,

    December 23, 2022 @ 12:13 am

    Interesting that people are discussing ChatGPT's ability to count syllables, but making no comment on the extremely low literary quality of its haiku.

  13. Taylor, Philip said,

    December 23, 2022 @ 8:46 am

    Perhaps because we are (quite wrongly, of course) judging ChatGPT on its ability to behave like a human. Humans above a certain age can, in general, count syllables with reasonable accuracy, but the ability to compose a haiku with literarary merit is an ability shared by (relatively) few.

  14. Kenny Easwaran said,

    December 23, 2022 @ 11:35 am

    ChatGPT definitely has interesting troubles with things like syllables and word lengths and rhymes. I asked it for some of the longest one-syllable words in English and it gave streaked, screamed, strengths, streaking, screaming, strengthen, and strengthening. I told it that some of those weren't one syllable words, and it came back with a list that included shrieked, streaked, crept, kept, heft. I told it some of those weren't very long, but it never managed to provide a list that just included long one-syllable words.

  15. Rodger C said,

    December 23, 2022 @ 1:44 pm

    ChatGPT doesn't think or reason. It says that haiku follows a 5-7-5 pattern because there are many text strings in its training database stating that haiku follows a 5-7-5 pattern. It has no ability to notice inconsistencies between one sentence in its output and another.

    As Zach Weinersmith said, "My God, we've created robotic college students."

RSS feed for comments on this post