Skip to content

3 Unsupervised Stanzas

June 2, 2010

I – default evaluation

one event unto every right work of
the living know not hasty in thy labour and money
all the rivers come to thee to all the
this that the world in the grasshopper shall the house also

II – eval: plosives, back vowels, assonance, alliteration

for all go unto god which hath
because to go to the ruler rise up
for there is good and to the peculiar treasure
behold that in his portion for man hath

III – eval: plosives, back vowels, assonance, alliteration (larger search)

again according to plant and to pluck up
to the clouds return to him power to the preacher
this sore travail and orchards and orchards and gone from her
better for the work of all sorts better

– April 16, 2010, unsupervised generation from type-based n-grams using stochastic beam search and phonemic evaluation.  Source text: Ecclesiastes (King James.)  Generator: ePoGeeS

The other day I posted something like: “I will summarize my approach to computer-generated poetry as: supervised poetry generation, with an eye to exploring and possibly answering questions in unsupervised poetry generation.”  So that’s kind of pretentious, OK, but from the start I assumed that the programming I did was a key part of the artistic endeavor.  One way to measure advances in the programming tools I’m developing is by seeing how well the programs do without me.

So I dropped in a text to build a model, set the generator to produce 4 lines, and clicked go.  The first time I just used the default generation (which I set just cause it looked good that way lol.)  The second time I minimized everything except for all the plosives and vowels (set to 1) and assonance and alliteration (which I maximized.)  The third time I used the second setting, but did a larger search (population of 30, 50 iterations.)  In all cases I took the first poems that came out; I didn’t reject poems for being bad, since I was hoping they’d be bad in interesting ways, which would be good.

But obviously the poems are bad in ways that reflect the limitations of the algorithmic approach as well as of my purpose, which was to develop a tool for me to use to write poetry.
– there’s a lack of rhythm or sonic structure, even with the phonemic selections of poems II and III.
– syntactic incoherence: lack of coherence due to the lines being generated individually, with a constant length defined by number of accented vowels.  In a supervised setting this is OK, ’cause you can just put the cursor to the end of the line and have it generate new words one at a time til it ends nicely.  (or you could do enjambment as in Gnoetry, and continue generating on the next line, though that would require having punctuation, which I wanted to do without for the moment.)
– discourse incoherence: the poems don’t have any unity or coherent narrative, even at the level of repeated imagery or any other type of poetic device.
– pragmatic meaninglessness: the poems don’t really mean anything at the pragmatic level; they’re not tied to any kind of agent who is referring to real-world experiences or objects, or coordinating its meaning with a population of individuals, which might include humans for example.

The syntactic and discourse incoherence are probably best addressed with some kind of grammar-based system.  I think that’s what Carpenter was using, though I’d prefer to have the grammar learned from a corpus.  I’d probably need some kind of semantics, though it’d be tough to learn that from corpus and tie it to the grammar in an automated way; probably some kind of authoring tool (maybe with suggestions learned from the corpus) would be the way to go.  There’s research in computational narrative that’d probably help out with discourse coherence.

Pragmatic meaning is interesting.  I’d like to make some kind of intelligent affective agent that travels a virtual world, and develops meaningful symbol sets tied to objects in the virtual world and coordinated with terms that it learns from humans that it meets.  The agent would then write haiku based on its experiences and locations its seen, and present them to the human who shared the environment with it.

So much fun and so little time… For now I think I’ll just focus on messing with the rhythm and sound, seeing what happens.

No comments yet

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: