Links to the implementations of neural conversational models for different frameworks. Contributions are welcomed.
Theano, Lasagne
A dialog system that is able to express emotions in a text conversation. See online demo.
Sample results for neutral emotions:
hi → hey you
how are you ? → i ' m fine , what about you ?
how old are you ? → i ' m 16 , how old are you ?
are you a leader or a follower ? → i ' m a man who is a politician .
my name is david . what is my name ? → i love your name .
is the sky blue or black ? → the sky is blue
you love anime . → i ' m not a weeb
Tensorflow
Nice results:
Hi → Hi.
What is your name ? → Laura.
What does that mean ? → I dunno.
How old are you ? → thirty-five.
Who is Laura ? → My brother.
Say 'goodbye' → Alright.
Two plus two → manny...
Modified code of machine translation model. No answering randomisation is implemented in this code, so the models answers with the same phrase way each time:
hello baby → hello
how old are you ? → twenty .
i am lonely → i am not
nice → you ' re not going to be okay .
so rude → i ' m sorry .
are you a robot or human? → no .
are you better than siri? → yes .
Torch
Source code for a bunch of super nice articles by Jiwei Li - one of my favorite researches in the domain of neural conversational systems. Worth checking out.
tell me ... how long have you had this falling sickness ? → a few months, I guess .
so i had the doctors test sammy ’s response to conditioning . → so he took the pills .
what are you going to do ? → i ’m going to get you a new car .
they fear your power your intellect . → i ’m not afraid of your power .
Hello? → Hi.
How are you? → I'm fine.
What's your name? → It's hard to describe.
How so? → I'm not sure.
What color is the sky? → It's blue.
What is your job? → It's not that i'm a fucking werewolf!
Keras
The most popular implementation of Seq2seq architecture on GitHub. However, there is still no evidence of getting good results with this repo.
A wrapper for farizrahman4u/seq2seq used for running experiments. No good results were achieved so far.
Need to check this up.
A new generative chatbot whose training converges in few epochs, including a model pre-trained on a small but consistent dataset collected from dialogues of English courses online.
user: Hello! How are you?
computer: i am fine , ok , well , thank you . (with probability of 0.540239)
user: What's your name?
computer: janet . (with probability of 0.319482)
user: What do you like to do now?
computer: let me think a minute . (with probability of 0.286760)
user: I want to chat about computer science.
computer: i am glad to hear there . (with probability of 0.293063)
user: Are you a real girl or a computer program?
computer: i am self employed . i would love to be well . (with probability of 0.026870)
user: Why don't you look for a job?
computer: i do not know . i am the bank . (with probability of 0.053494)
user: So, you don't need a job!
computer: no . i have to make a lot of money . (with probability of 0.001279)
A great overview of existing conversational datasets.
Get a lot of raw movie subtitles (~1.2Gb)
~ 40Mb after clearing out the technical data.
Refer to this great review by Richard Krisztian Csaky: Chatbot and Related Research Paper Notes with Images
as well as to his report paper in pdf: Deep Learning Based Chatbot Models