5 ms·

Has anyone tried to make a poetry generator that has some semblance of thematic unity ? Otherwise you just get a word salad generator that produces good poems

Has anyone tried to make a poetry generator that has some semblance of thematic unity ? Otherwise you just get a word salad generator that produces good poems by accident. Maybe we need to try word2vec based approaches.

pizza 10y ago
Devil's advocate: the thematic unity is a type of beauty that lies in the reader's understanding ;) To be fair, it does seem like this poem almost has some nice pseudo-rhyming scheme in it
- thewarrior 10y ago
  I just don't find them interesting as a poet. Even a markov chain can produce good poems once in a while. This is impressive from an AI and machine learning perspective but NLP is one area where I feel there haven't been any fundamental breakthroughs yet as we've seen for images. Perhaps language is much more complex. Visual processing is present in many animals and so it might be that it's simpler at some level. There are very few animals that have evolved to use language like we do. So the underlying models must be far more complex than stacking layers of convolutional layers on top of each other. And compared to the spectacular results in images like Neural style transfer , we're judging these to a very low standard. Here is an actual poem : “You know how this is: if I look at the crystal moon, at the red branch of the slow autumn at my window, if I touch near the fire the impalpable ash or the wrinkled body of the log, everything carries me to you, as if everything that exists, aromas, light, metals, were little boats that sail toward those isles of yours that wait for me.” ― Pablo Neruda, The Captain's Verses
  - Dzugaru 10y ago
    I bet images like Neural style transfer are boring to artists just like ML poems :) Actual digital art is no less complex [0]. That aside, it would be interesting to look at ML model used, all I've seen were simple ones like hidden markov chains for the next letter, but hidden state is managed by RNN. You need to learn to "grasp" some underlying meaning if you want more than a semi-random word salad and if this can actually be done without building a strong AI in the process :) [0] https://www.artstation.com/artwork/0ogb8 https://www.artstation.com/artwork/0ogb8
sdenton4 10y ago
Yes, of course. Anyone who follows the field knows that Trurl built a most amazing electronic bard, but the end result was a bunch of dead poets and supernovae. So these days we stick to the word salad approach, to avoid repeating that calamity.
- pacaro 10y ago
  Just in case you are missing the reference, this is from Stanislaw Lem' "Cyberiad" This particular sally can be found at http://sfbay-anarchists.org/wp-content/uploads/2012/05/Trurls-Electronic-Bard.pdf http://sfbay-anarchists.org/wp-content/uploads/2012/05/Trurl...
puppymaster 10y ago
Not unity per se, but I did put together one using Backus-Naur technique https://poem.alv.in/ https://poem.alv.in/
- thewarrior 10y ago
  This is actually much more interesting than the OP. What kind of grammar did you design ? Do you think we can machine learn the grammar aided by things like word2vec ?

pizza 10y ago

Devil's advocate: the thematic unity is a type of beauty that lies in the reader's understanding ;) To be fair, it does seem like this poem almost has some nice pseudo-rhyming scheme in it

thewarrior 10y ago
I just don't find them interesting as a poet. Even a markov chain can produce good poems once in a while. This is impressive from an AI and machine learning perspective but NLP is one area where I feel there haven't been any fundamental breakthroughs yet as we've seen for images. Perhaps language is much more complex. Visual processing is present in many animals and so it might be that it's simpler at some level. There are very few animals that have evolved to use language like we do. So the underlying models must be far more complex than stacking layers of convolutional layers on top of each other. And compared to the spectacular results in images like Neural style transfer , we're judging these to a very low standard. Here is an actual poem : “You know how this is: if I look at the crystal moon, at the red branch of the slow autumn at my window, if I touch near the fire the impalpable ash or the wrinkled body of the log, everything carries me to you, as if everything that exists, aromas, light, metals, were little boats that sail toward those isles of yours that wait for me.” ― Pablo Neruda, The Captain's Verses
- Dzugaru 10y ago
  I bet images like Neural style transfer are boring to artists just like ML poems :) Actual digital art is no less complex [0]. That aside, it would be interesting to look at ML model used, all I've seen were simple ones like hidden markov chains for the next letter, but hidden state is managed by RNN. You need to learn to "grasp" some underlying meaning if you want more than a semi-random word salad and if this can actually be done without building a strong AI in the process :) [0] https://www.artstation.com/artwork/0ogb8 https://www.artstation.com/artwork/0ogb8

sdenton4 10y ago

Yes, of course. Anyone who follows the field knows that Trurl built a most amazing electronic bard, but the end result was a bunch of dead poets and supernovae. So these days we stick to the word salad approach, to avoid repeating that calamity.

pacaro 10y ago
Just in case you are missing the reference, this is from Stanislaw Lem' "Cyberiad" This particular sally can be found at http://sfbay-anarchists.org/wp-content/uploads/2012/05/Trurls-Electronic-Bard.pdf http://sfbay-anarchists.org/wp-content/uploads/2012/05/Trurl...

puppymaster 10y ago

Not unity per se, but I did put together one using Backus-Naur technique https://poem.alv.in/ https://poem.alv.in/

thewarrior 10y ago
This is actually much more interesting than the OP. What kind of grammar did you design ? Do you think we can machine learn the grammar aided by things like word2vec ?