Here is a slightly more sophisticated word based algorithm written in Scala, that takes the probabilities of word sequences of arbitrary length into account. (That's not the original dissociated press algorithm.)
The Algorithm is as follows. In each step select a rolling half of the text starting at a random position, search for the longest tail sequence of output words that occurs in that half (this might be 0 words) and output the next word.
import io._, collection.mutable.ArrayBuffer, util.Random
import java.io.FileInputStream
val lines = new BufferedSource(new FileInputStream("markov.txt")) getLines
val wordregex = "\\b[a-zA-Z]+\\b|[.,?!]".r
val words = lines flatMap (wordregex findAllIn _) toArray
val rollingwords = words ++ words.slice(0, words.length / 2)
val rnd = new Random()
val outwords = new ArrayBuffer[String]()
for (i <- 1 to 1000) {
val startposition = rnd nextInt (words.length * 2 / 3)
val half = rollingwords slice (startposition, startposition + words.length / 3)
var newword = ""; var n = 0; var index = 0
while (index >= 0 && n < half.length && n < outwords.length) {
index = half.indexOfSlice(outwords.takeRight(n))
if (index >= 0 && index < half.length - n) {
newword = half(index + n)
}
n = n + 1
}
outwords += newword
}
println(outwords.foldLeft("")(_ + " " + _))
Here is a sample output also generated from the wikipedia article on markov chains:
today stationary distributions will not be unique I probabilities
satisfy k rightarrow position non the transition probability
distribution can be represented mapping only if the parameters on the
unit the system , Allowing n to be unique , in that i in the
stationary distribution or invariant measure if it satisfies the
stationary distribution for Q .
By the way, if you use "[a-zA-Z .,!?]".r as wordregex you can use this to generate letter based dissociated press as well:
This figurrent or periods when a backgrobability the Pater ext state
with stochare a number of detelemely if theresting class of where Mi
pimatransie, opens that the nnn need need by a system state is
errords, then limpor all task.
It gets really interesting with a large text file like the Jargon file. Now letter based is already quite good:
Other direction algorithm will happily errors, and an uncommon; it's
been shorthand for "out being proms, and meta-location hack with
decades built around the LISP Mac pre-Internet access workstation.
This may be dead. A measure of competitors, a popular compiler end
repeatedly to second, and was leech. arranged with the encountered on
the net, esp. from a network. Usually `customer and on the chad it
back onto paper. Several had in the unique properties.
Wordbased becomes quite amusing:
This has since been reported . The only thing it expects one resource
leak n . A semi - mythical language construct in an inconsistent
because it can t adjust in the first place . If you enter a computer
in a playful and ended Get a real computer ! imp . Sarcastic
invitation to say Talking . Small cable were blamed for real
programming . Pascal ten years later , but the majority of our product
not quite the same modern subshell . There is some dispute over
whether this entry everybody s mother .
