NaNoGenMo
Entry: Towers in the Land of Smiling Chimeras
I'm new to this, but excited to try it out!
My idea for now is to do a progressively lipogrammatic novel a la Ella Minnow Pea: so the letters that are "allowed" to be used decrease as the novel goes on. Currently I have a basic outline of how to do this that may change, but here's the general idea:
- generate a 50K text (using Markov chains at least to begin with), divided into chapters (each chapter will lose an additional letter eventually, but ignore that constraint in this step)
- go through each chapter to find words that use the forbidden letters. in order,
- replace some fraction (yet to be determined) of banned words with synonyms (probably with WordNet), hoping that this leads to increasing vocabulary contrivedness as characters/the narrator try to get around the letter-based constraints.
- replace the rest of banned words (either ignored in the first step or things like prepositions, etc that would be hard to find synonyms for) with letters that look/sound similar to banned ones - e.g. Ella Minnow Pea uses "et" instead of "ed" for the past tense - "walket" instead of "walked". This part will probably be the hardest, because I'd like to play with sounds in some places but English vowels don't like to cooperate with that.
- black out/redact any words that remain - for which there weren't synonyms or there weren't good letter replacements - ideally the blacked-out portion of the text will increase as the novel goes on.
I imagine I'll have to toy with the ratios and letter replacements above to get something that's a balance of readable and going with the concept. I'm also not totally sure yet what type of texts I want to use to train my Markov model on - I think it'd be fun to do political/spy novels or something like that to play with the redacted elements, but I don't know if I'd be able to find enough text to get a good model.
Possible future improvements, once I get a basic model working:
- Switch from Markov chains to some kind of neural net for generation of the initial text - I'm not sure about how these work but I'm interested in them, so this might be a good opportunity to try
- try to improve the quality of the generated text - differentiating beginning from end, keeping consistent characters, things like that
- look into "invented synonyms" - EMP does stuff like "aeiouy" for "vowel" and "grapheme" for "letter" that WordNet might not have, but that add something to the text.
- playing around with the presentation/design, maybe adding poem-based epigrams to chapter headings
Some gems from toying around with Markov generation, using Pride and Prejudice and Moby Dick because they were convenient:
"I was sure you could not be married all day" "The envelope contained a sheet of blubber." "Hold the steak in one hand, and a still slighter shuffling of women's shoes, and all was soon right again."
Can't wait to see what happens when I get to synonym substitution
Well, I've got a pretty decent corpus of mystery novels that's giving me the tone I want (or at least a good enough approximation - I particularly love the last sentece) - on to synonym substitution!
Also, I think Markov chains are better at using semicolons in writing than I am.
His dinner cost ninety francs; he threw his last hundred-franc note on the table, and sat down at his desk, and rapidly wrote two letters.I have speculated on the chance that Mr. Luker is lying--and you assume again that the Shivering Sand itself smiling at him in astonishment.Perhaps nothing in the world.The one interpretation that I could discover was altered in her pretty summer dress.Penelope resented my merriment, in rather a headlong manner into talk with strangers, without waiting to feel his continued restlessness at night.He was seized with a violent, persistent, tenacious craving for pepper.
Synonym replacement is working! (After a whole lot of fighting with nltk and various encoding errors)
As a celebration, my favorite sentence from the most recent test run:
He firmly repeated his intention to retire in spite of a general disbelief in the existence of any lunar atmosphere at all
It's done! Or at least has reached 50K - I might yet make some improvements, but all the basic things - synonym generation, misspelling, and redaction - are working. Code is here, novel is here. I intend to do better documentation before November ends and maybe also make some improvements, but for a busy November, I'm happy with the way it turned out.
Changing title to my favorite phrase from the generated novel: Towers in the Land of Smiling Chimeras
