4 ms·

In the article on your blog, you wrote: "GPT-3 rhymes reasonably well and often when appropriate, but the improvement is much smaller on rhyming than it is on

• burnished 3y ago
  That seems incredibly challenging, I'd expect some fundamental difficulty due to rhyming being determined by how a word sounds and not what it means.
• gwern 3y ago
  > However, once they've started writing rhyming poetry it's hard to get them to stop rhyming. They seem to have formed a strong association between rhyming and poetry. I've also been unable to get them to obey a specific rhyming scheme like ABBAB. Correct and commonly observed (eg. https://arxiv.org/abs/2305.11064 https://arxiv.org/abs/2305.11064 ). (At least, for GPT models. I don't know as much about the Anthropic models as I should, although I understand they do still use a BPE tokenization, unfortunately.) My theory is that it is a surprising interaction of BPEs with RLHF: https://www.lesswrong.com/posts/t9svvNPNmFf5Qa3TA/mysteries-of-mode-collapse?commentId=tHhsnntni7WHFzR3x https://www.lesswrong.com/posts/t9svvNPNmFf5Qa3TA/mysteries-... As much as they look like they can, they can't rhyme because of BPEs still. What they have done in lieu of genuine phonetic understanding is, more or less, memorized a ton of rhyme-pairs: they only have a vast patchwork of half-understood phonetics discerned dimly through the lossy compression of BPEs and memorized pairs. If you don't force them out of the memorized space and let them write without interruption, they look like they understand, but they still don't. Then RLHF punishes them for any incorrect poetry, so they never leave the memorized space on their own because that's the only way to guarantee correct rhyming poetry. And since there is no way for it to tell the difference between 'rhymes but I don't know that it rhymes because BPEs' and 'deliberately nonrhyming poetry', much less what the difference is between 'ABBAB' and 'AABBAA', it just always does rhyming quatrains etc. Why take the risk? Also applies to jokes and joke explanations: https://arxiv.org/abs/2306.04563 https://arxiv.org/abs/2306.04563 It can't understand properly what is a joke or not, because it's blind to what makes a vast number of jokes work, so it just memorizes a few safe jokes and assumes anything presented to it as a joke must be one of the countless jokes that it can't understand & makes up its best guess.
  • zarzavat 3y ago
    I wonder if having access to characters actually helps rhyming in English all that much, as English rules of pronunciation are essentially rote-learned anyway. If it were not rote-learning, then it might make different mistakes, for example expecting two words to rhyme because they end with the same suffix. Perhaps it would be more effective to ask it to produce poems in the format: English0 IPA0 English1 IPA1, where each line is produced in both semantic and phonetic representations. This would give it the context necessary to “see” the rhymes without having to mess around with the tokenization.
    • telotortium 3y ago
      Can GPT accurately transcribe English text to IPA? It looks like it can (https://chat.openai.com/c/c893cf54-5dfa-4774-9800-e6a74b7f9e2e https://chat.openai.com/c/c893cf54-5dfa-4774-9800-e6a74b7f9e...): Prompt: Eng Rhyming with IPA User Transcribe the following English text to IPA: English: I wonder if having access to characters actually helps rhyming in English all that much, as English rules of pronunciation are essentially rote-learned anyway. If it were not rote-learning, then it might make different mistakes, for example expecting two words to rhyme because they end with the same suffix. Perhaps it would be more effective to ask it to produce poems in the format: English0 IPA0 English1 IPA1, where each line is produced in both semantic and phonetic representations. This would give it the context necessary to “see” the rhymes without having to mess around with the tokenization. IPA: Response: aɪ ˈwʌndər ɪf ˈhævɪŋ ˈæksɛs tuː ˈkærəktərz ˈækʧuəli hɛlps ˈraɪmɪŋ ɪn ˈɪŋglɪʃ ɔːl ðæt mʌtʃ, æz ˈɪŋglɪʃ ruːlz ʌv prəˌnʌnsiˈeɪʃən ɑːr ɪˈsɛnʃəli roʊt-lɜrnd ˈɛniˌweɪ. ˈɪf ɪt wər nɑt roʊt-ˈlɜrnɪŋ, ðɛn ɪt maɪt meɪk ˈdɪfərənt mɪsˈteɪks, fɔr ɪgˈzæmpl̩ ɪkˈspɛktɪŋ tuː wɜrdz tuː raɪm bɪˈkɔz ðeɪ ɛnd wɪð ðə seɪm ˈsʌfɪks. pərˈhæps ɪt wʊd biː mɔr ɪˈfɛktɪv tuː ˈæsk ɪt tuː prəˈdus poʊəmz ɪn ðə ˈfɔrmæt: ˈɪŋglɪʃ0 ˈaɪpiːˈeɪ0 ˈɪŋglɪʃ1 ˈaɪpiːˈeɪ1, wɛr iʧ laɪn ɪz prəˈdusd ɪn boʊθ sɪˈmæntɪk ænd fəˈnɛtɪk ˌrɛprɪzɛnˈteɪʃənz. ðɪs wʊd gɪv ɪt ðə ˈkɑntɛkst ˈnɛsəˌsɛri tuː si ðə raɪmz wɪˈðaʊt ˈhævɪŋ tuː mɛs ɚˈaʊnd wɪð ðə ˌtoʊkənaɪˈzeɪʃən.
    • gwern 3y ago
      I did some poking around with IPA way back in 2020 reasoning that if the phonetics were explicit maybe that'd be fine, but I didn't get anything that looked like a big improvement: https://gwern.net/gpt-3#ipa-rhyme-annotations https://gwern.net/gpt-3#ipa-rhyme-annotations My guess was that it doesn't see enough IPA to use it that way zero-shot, and the erasure of information by BPEs damages its learning fundamentally enough that you can't easily prompt your way into anything better. I speculated that because it's memorized so much, it shouldn't be too hard for it to learn to rhyme properly if you finetuned it on an IPA-encoded or non-BPE-tokenized poetry corpus, but I never got around to it and AFAIK no one else has tried that yet.