Fractal Zero Shot Chatbot


Author: Tofara Moyo

Title: Fractal Zero Shot Chatbot…/332330414_FRACTAL_BASED_ZERO…

We outline a way to exploit the fractal nature of
conversation in order to predict the next word given
all the preceding ones in the generation process in
a zero shot trained chatbot. First a corpus of
conversations is given. Then for every instance of a
word, we create a set that contains an unordered
list of all the words that follow after it. We have
unique entries for each unique instance of a word that came after
in each set. And give each set a value. At
initialisation this will be zero. During operation when
given a particular word we increase the value of
every set that that word is associated with. Then we
take the entire database of sets and choose one
element from it at random. That word will be selected as the
next word in the utterance. The probability of each word in the sets is increaded proportionaly to their value.The process is repeated
with this new word, but the value we gave to the
previous sets is maintained. We sample yet again
from the union of all the sets to select the next word. This process is
continued until an end of utterance symbol is met.
When a human agent responds each of the sets
corresponding with the words uttered get their value
incremented, when the utterance is complete, the
chat bot will know the first word to say, and the next
and so on. This chatbot can keep cognisance of all that has gone on in a conversation to the current point, regardless of the number of exchanges. Because of the fractal nature of
conversation we shall see that this should work.



If you did this with a human, making a set of all neurons that fired after a particular one be the basis , and mapping it to inputs and outputs that we also create sets for, you could have a zero shot trained robot.