>> Does this actually work? Okay. Cool.
Um... We're good? We're starting? And
we're up there. Okay. Well, hi. My name
is Danielle and this Darius, and I'm up
here because I really love poetry. I
know this is a conference for people who
love coding. I'm into that too. Don't
get me wrong. But let's focus on poetry.
Darius and I are doing this together
because we separately wrote some
programs where we found or created
poetry, and we found some similar things
along the way. I guess I'm supposed to
do this one. So we did stuff with
accidental poetry, which is poetry
that's either detected in or generated
from existing texts. My project was
called Nantucket, and I looked for
limericks in books. You look for rhyming
and syllable count and see it happens to
be a limerick. Darius did Markov chain
stuff where he created poetry. Which was
better.

 >> (inaudible) gets more and more
interesting results. You don't need
fancy theory.

 >> So this is an example of one of the
accidental limericks that I found. This
is in Helen Keller's autobiography, the
story of my life, where she happened to
write... (on screen)

 >> This is from a Markov chain from a
19th century cookbook.

(on screen)

 >> So we both started with something
like this. Take the CMU pronouncing
dictionary, look words up in it, and
check against some approximation of
meter and rhyme. So here's the
dictionary. You get pronunciations, one
per line, and here are two of the lines.
Pronunciations for the same word, on the
left. On the right, the way to say it,
two different phonemes. Ah-dress and
ah-DRESS. You can see the different
stress. AH-dress and A-DRESS. So
naturally it goes like... Reclaimer,
spare that tree. With this, we get
reclaim er spare that tree. So start.

 >> So we have to deal with a lot of
words that aren't in the CMU dictionary.
So you have some choices. How do you
deal with that? You can ignore those
words and only pay attention to words
you find in the dictionary, which is
kind of boring. You miss out on a lot.
You can do really complicated stuff, and
that sounds fun, but sometimes you want
to play, or you can have hacky ways of
guessing pronunciation. You can guess
which one of us did. I played around
with finding rhyming in words that
weren't in the CMU dictionary. Even if
the word wasn't in the dictionary, maybe
the last syllable of that word was in
words that were in the dictionary. So I
built my own pronunciation dictionary
that was just last syllables of words.
So I had to find the last syllable worth
of characters in a string. This is
horrible. I wrote it a few years ago,
and this was my first Python, so I
apologize. But I wrote this crazy
regular expression. I swear, I haven't
updated it in years. But the basic idea
is I'm looking for one or two
consonants, followed by any number of
vowels, followed by optionally some
conscious
es, maybe e or ed, because those don't
add syllables at the end of the word.
Let's move on and never look at that
again. So you can find the last syllable
worth of phonemes to follow with that
sound. So I just have to look at the
vowel sounds. We know it's a vowel sound
because it ends in a digit or stress as
Darius was saying. So I find a digit at
the end. Anything after that, great. So
that let me find -- words like this,
from the insect world by a guy whose
last name I can't pronounce. The
chrysalis or pupa the perfect insect
orthoptera... (on screen)

And also, just one more from the King
James bible.

(on screen)

So... Darius also played around with the
King James bible.

 >> (on screen)

 >> So technically speaking, rhyming
depends on more than just the
pronunciation of the last syllable of
each word. You're supposed to pay
attention to the last stressed syllable,
not use repeated words, but if I limited
myself too much I wasn't going to get as
much fun. So I have bad rhymes, but I
get these cool limericks like from
Dracula. (on screen)

 >> Okay. Meter in verse means a regular
pattern of beats and here we mark the on
and off beats. Reclaimer, spare that
tree. The pattern is not there in the
CMUdic but we get the right amount of
syllables by counting the vowel sounds
and some of these ones and zeroes
correspond to on is off beats, one for
on, zero for off. Some of them are on
off beats. We could say -- a one can
match an offbeat if it's the only
syllable. But that's too permissive. It
would say that this is perfect iambic
pentameter. (on screen)

Most of it sounds almost that bad.
Ultimately we want fancy machine
learning, and honestly, that would have
made way more sense than my little pile
of hacks, but there's an addictive
quality to doing what you can and seeing
what comes out. (inaudible) Lord of the
Rings. (on screen)

 >> So another thing we both had to deal
with was counting syllables. Again, it's
super easy for words in the CMU
dictionary. Syllable -- just count the
number of vowel sounds. They end in
integers, count the number of phonemes
that end in integers. Great. Four
syllables in concatenate. What about
those not in the CMU dict? You can count
the number of vowels, but it wouldn't
work for room and board, that have
multiple vowels that make one vowel
sound. We could count the contiguous
groupings of vowels. But this has
problems with words like uglier and
contiguous. Where you have groupings of
vowels with multiple vowel sounds. So
what do you do? We counted the number of
contiguous groupings of vowel sounds and
(inaudible) for each digraph. Pairs of
vowels that usually form a single vowel
sound. A little fuzzy, going to screw it
up sometimes, but good enough for
government work. There are other things
we messed around with, like ious is one
vowel sound. Or maybe even an
apostrophe. That adds a vowel found.
This is a really short talk. So ignoring
various restrictions and really giving
ourselves permission to have bad rhymes,
bad meter, something that's close
enough, let's get a lot more interesting
stuff. Especially looking for accidental
verse, where I have a very limited
resource. I didn't reject rhymes that
use exactly the same word, so I got ones
like this one, from the Count of Monte
Cristo.

(on screen)

And I also didn't require that I have an
AABBA rhyme scheme. I was okay if every
single line rhymed. Which is a terrible
limerick. But it helped me get these
from the Just So Stories.

(on screen)

I'm really okay with anything that lets
me say frouzly in front of all of you.
And one last fabulous one from Darius.

(on screen)

 >> Thank you.

(applause)