A ::SHE+ILA:: AI Chatbot [$80]

More about A ::SHE+ILA:: AI Chatbot [$80]

Unlike the procedural chatbots (such as wooeebot.dll, which used to write stories on my old Janet and Julia Story Forum (wandanet.co.uk -- no longer online)), an AI chatbot uses neural network technology to parse sample human language prose and produce realistic output. Training may be via backpropagation, driven by samples of real-world text (such as an http webscraper), or other feedback mechanisms which adjust the neural matrix weightings; or by the "weasel" method which generates random matrices and tests their output against a set of rules (the "weasel" app produced random atoms and compared the output prose with a target string, stickifying those atoms which correlated successfully againt the target). ::SHE+ILA::'s dictionary contains about 10,000 English words classified by part-of-speech (noun, adjective, verb, adverb, preposition, conjunction, etc.) in a huge ETAIONS-compressed Buf indexed by "atom" (an integer which represents the word throughout the system), and in the original Madlib-type test program generated random sentences like "As Fortran as the gibbon", and "Anette sat on the blue marxist chair eating strawberries", etc. The true AI version would start off by creating a number of randomized state machine Bufs, and adjust their contents to maximise the correlation of the output prose. It could be trained by feeding it samples of public-domain text documents, such as the Rubiyat and the Bhagavad Gita. Also the contents of several Usenet newsgroups (I have tons of these backed up, lol). Live web pages may tend to be politically biassed, so would not be used at first. A mapping space would be output, showing which atoms (words) follow each other in normal English usage, and maybe Markov chains to represent their relative probabilities. It would be interesting to see what sort of drawings DDG would produce if ::SHE+ILA::'s output were used as prompts also :)