https://www.researchgate.net/publication/332292331_ZERO_SHOT_CHATBOT
abstract
We outline a method for creating a dynamic Chatbot
without incorporating a training phase in the process. It is
first noted that even if a random sequence of numbers is
mapped to an “in sequence” range of alphabetical letters, a book written by substituting those randomly initialised numbers with the letters would still be easily readable by an experienced cryptographer. We then create an elaborate mapping of words to numbers, preceding words, found
in a conversational corpus ,mapped to the following words
numbers, and preceding numbers mapped to the following words .With a bit more elaboration in the pathways the Chatbot is ready to operate simply by following these paths. .
the Chatbot can handle topics
relatively largely divorced from the material found in the
conversational corpus.
Of course if we have multiple conversations to map we would have a subscript for each conversation"a" and “b” having within them the following statements somewhere
HALLO-1a ;HOW-1a ; ARE-2a, YOU-1a…
and the other one.
…HALLO-1b ;BUDDY-1b ;HOW-2b; YOU-1b ;DOING-2b;TODAY-1b.
…i can show how we could have a novel introduction…we are started of with the word HALLO…stripped of number and subscript…and the system will choose low values for the number and any letter…say it chooses 1a…so we have “HALLO-1a” then the 1 maps to “HOW-1” as it did above…but instead of a HOW -1 with an “a” subscript, we choose the subscript at random and say get a “b”…that means the sequence becomes …
.“HALLO-1a” ; “HOW-1b”…
…and as we can see above, “HOW - 1b” is mapped to “YOU-1b” ultimately we get the sentence
“HALLO”, “HOW”,“YOU”;“DOING” ; “TODAY”…a unique sentence from those found in the corpus
mapping words to numbers make you able ,with cryptography ,to decipher a conversation written with those numbers , and mapping those combined units to each other in a heterogenous way makes you able to decipher the rules for making conversations in general
This is a property of the fact that we found a way to order the elements in the converstion according to two criteria, increasing number according to instantiation position within the progresion of the corpus for words…these words got their numbers at specific points within the text and that was not random…it correlated with all the other words getting their numbers at specific points within the text because of the mapping all the properties of these points are captured in one still picture…those points are fully determined by the rules of discourse as used in the corpus’ according to their distribution…that is how mulitple contexts are captured and represented…and further alows for dynamic conversations
This is similar to how “suggest a word” functions in cell phones work.
Certain high frequency combinations and sequences are what is learned.
In more advanced models the structure of these sequences as also learned.
So what kind of technology is behind this widely available function?
HTM has inspired a champion in the language processing area:
https://www.cortical.io/
The core of going from just stringing common word combinations together to working with words that have a similar meaning:
What can you do with this technology?
This is simply an extension of my earlier model greatly revamped…there is no learning algorithm we just give each instance of a word a number starting with 1 for the first instance, then map the words cross wise to the numbers and the numbers to the words, between consecutive pairs in the forward direction…this is exactly the same as the Dynamic spiking network in functionality…the numbers represent the hidden theta nodes while the utterances are inputs and outputs represented as nodes whenever encountered or produced…
So if we have the sentence “the people are hungry” and “everyone is famished” i beleive the sytem could produce the second statement rather than the first in spite of being not trained on it directly, but only when each of the words in the second sentence have been used in all the contexts they exist in the wild, and perhaps the first was a reply to a particular sentence within the corpus…
well what we have is a spiking network with hidden neurons represented by the numbers…also we have trained the sytem to the level of completion once we do the mapping because the nodes before fully determine the nodes in the next stage just like the dynamic fuzzy spiking ones did…because we use only two nodes at any one time this is a sparesly distributed system given that of the total number only two “fire” at any one time…also we have the same number being used to map different outputs, and the same output only giving different outputs if the other nodes it is with are exactly the same…i.e. the exact situation we designed in the dynamic spiking net is here…what it is like is that the centroid 1 when firing with the centroid input “bye” will ,if it lead to the next nodes firing being 4 and ok…then these two “4” and “ok” will exist somewhere within the spheres of centroids 1 and bye =IN EFFECT=…if we then ,in a totaly different conversation have the sitution of centroid 9 and centroid bye firing, leading to nodes 3 and node chow firing after them we would in effect have 3 and chow within the radius of 9 and bye because of the cross heterogenous mappings “bye” will have both 4 and 9 within its radius but at different distances just right for the previous two examples to have happen to happen ,where we have the same input “bye” and one moment we say ok and another time we say chow in return…the more instances exiting of combinations of words and numbers the greater the number of nodes present within multiple raddii of centroids…This simmilarity with the dynamic spiking network is only an isomorphism…nowhere in memory have we stored free nodes within centroids…the sytem just acts as if it has and is calculating probabilities each time is fires nodes…the information is stored with in the explicit mappings…so this will submit to the same reasoning i used to estblish that similar inputs lead to similar outputs in the dynamic neural network.
Look at the Cortical IO videos.
You should recognize them as a greatly advanced version of what you propose.
Instead of numbers they use bits and topographic location - the same thing in a new suit.
1 Like
I will have a look when i have a bit more time , thankyou for them!