I have thought about this quite a bit and I believe that this particular pairing, Nietzsche with Family Circus, is at least especially good (if not uniquely good).
For me, Family Circus was never really enjoyable. Somehow that makes Nietzsche Family Circus even better.
My guess is that if you start pairing things at random, you will on average going to get something much, much worse than this. An understanding of why this is true promises to elucidate the contemporary condition, I suspect.
That's one of the fun little oddities Gary Larson talks about in "The Prehistory of the Far Side", where syndication captioning errors led to a couple of swaps of Far Side and Dennis the Menace cartoons, with weird results.
Maybe the language model could use some smoothing? It seems to frequently regurgitate lines from existing comics rather than composing new ones. For example, [1] has "Either mom's cooking dinner, or somebody got sick in the kitchen..." in panel 3, which is almost entirely the 1 June 1986 comic [2] which has "Either mom's cooking dinner, or somebody got sick in the furnace duct."
Cool work nonetheless, and a fun way to re-encounter a favorite.
I'd wager it's due to the limited corpus currently entered, plus the presumably limited corpus of the entire comic overall. How often does something else follow "Either mom's", "mom's cooking", "cooking dinner" other than the following words in this?
(I haven't looked at the actual implementation, so I dunno how the actual stuff is generated.)
Your wager is good, on both counts; the current corpus is based on the first 14 months of strips, which comes out to about 10K total words for Calvin (who has more lines than the rest of the cast combined, in that chunk of strips). That's not enough to generate a lot of variety on anything other than very common word combinations. Strings of prepositions and articles are the most likely inflection points where you'll commonly see two distinct phrases glued together.
If I get the whole strip run into the corpus, that'll kick Calvin up to something more like 80-90K words, which will help with the variety a good bit, and the other characters will have more of a shot of current Calvin-like variety, but it's still a relatively small training set.
By comparison I've done some markov model experiments based on multi-million word corpora and that gets a lot farther into the territory of regularly producing satisfyingly weird disjunctures.
13 comments
[ 2.2 ms ] story [ 29.5 ms ] threadFor me, Family Circus was never really enjoyable. Somehow that makes Nietzsche Family Circus even better.
My guess is that if you start pairing things at random, you will on average going to get something much, much worse than this. An understanding of why this is true promises to elucidate the contemporary condition, I suspect.
http://www.joshmillard.com/markov/calvin/?215934307460273100...
http://www.joshmillard.com/2015/07/06/calvin-and-markov/
Cool work nonetheless, and a fun way to re-encounter a favorite.
[1]: http://www.joshmillard.com/markov/calvin/?657026408452153958... [2]: http://www.gocomics.com/calvinandhobbes/1986/06/01
(I haven't looked at the actual implementation, so I dunno how the actual stuff is generated.)
If I get the whole strip run into the corpus, that'll kick Calvin up to something more like 80-90K words, which will help with the variety a good bit, and the other characters will have more of a shot of current Calvin-like variety, but it's still a relatively small training set.
By comparison I've done some markov model experiments based on multi-million word corpora and that gets a lot farther into the territory of regularly producing satisfyingly weird disjunctures.