not an issue - GAN vs chatterbot / ai bot architecture

[johndpope]

I threw this code together which launches the chatterbot code directly against your trained model.
pip install git+git://github.com/gunthercox/ChatterBot.git

https://gist.github.com/johndpope/b0d9a025c6e54dc1e07ab6100c34a24a

the results were quite bad.

there's a bunch of yml files in chatterbot which if I get time will attempt to glue together - but it would be good to be able to have an api into this architecture where it could be corrected...or maybe that exists in codebase?

I understand this repo is to showcase whitepaper research - but wondered if you have ideas to take ai bots to another level. If you were to step back - and look at higher level of what could be achieved with blending technologies.

off the back of an envelope - some thoughts of mine
eg. by doing entity detection on sentence / have these carry forward
memory of stuff (I guess this opens up problems like microsoft's disastrous tai tai twitter bot)
eg. use elmo for understanding context of words
wikidata -> don't stop at entity detection / go beyond to understand meaning sentence.
use hybrid code network / action templates
Commonsense Knowledge

what do you see as the biggest challenge that needs solving?

[oswaldoludwig]

First point: This isn't the best GAN-based model, the best one is here. This version uses the discriminator of the GAN-based model to select the best answer between two models, one trained by teacher forcing and another trained by our new GAN-like training method.
Second point: you're comparing two different technologies, a retrieval model with a generative model. The retrieval model isn't able to create new answers, it only retrieves an answer from a set of answers, provided by the user. The generative model learns to create new answers by "reading" a large amount of text. Generative models seem to be the future, but they're still inside the research labs, since they're much more complex. If I had ideas to take generative bots to another level, I would be a very rich man :-).

[oswaldoludwig]

I think one of the biggest challenges that needs to be solved is the acquisition of background knowledge.

[johndpope]

did you consider plugging in something like concept net to the conversation pipeline ?
https://github.com/commonsense/conceptnet5

https://arxiv.org/pdf/1805.07858.pdf

[oswaldoludwig]

This is a relevant topic, thanks for the link!

[johndpope]

so been digging around and came across sling parser framework by google.
sample code is thin -
https://github.com/google/sling

came across this repo
https://github.com/sih4sing5hong5/sling_trying_out/blob/df3604419a6849dba1df2a6b8779a9eef1799bdf/trying_out.py

import sling

parser = sling.Parser("sempar.flow")

text = raw_input("text: ")
doc = parser.parse(text)
print doc.frame.data(pretty=True)
for m in doc.mentions:
  print "mention", doc.phrase(m.begin, m.end)

naively speculating, if the word2vec embeddings in this agent could incorporate these frames / tokens - which include a transition system to detection actions within text - evocations / shift / refer / assign / embed / elaborate - wouldn't this help propel understanding in latent space of end to end training?

eg. there's some images here
https://ai.googleblog.com/2017/11/sling-natural-language-frame-semantic.html

white paper here
https://arxiv.org/pdf/1710.07032.pdf

eg. john hit the ball.