i retired from my downward lascerations on my
wrists to die." computers do most
of the altar. the soundtrack album
basic form, the
test assumes that there
are no parameters to
next series of
boats sylphe. on the money, high-powered
jeff became close to the
the motion of an object. in
I'm thinking of upgrading the wikipedia-poetry script from a set of randomized Markov chains to a recurrent neural network. I'd prefer to see output that toes the line between interesting and coherent but the current approach leans too far toward the former. This blog post should be helpful - see the comments for a fun example using generated NSF grant abstracts.
Of course, the most obvious use of an RNN here would be to generate wikipedia entries. They could still be poetic if we cut and paste bits of them together at meaningful intervals. Perhaps the RNN could be taught to recognize the emotional value of certain words (i.e., the words die or money in the above example have some emotional impact, while jeff might offer comparatively little).