This is a chatbot trained by seq2seq and reinforcement learning.
- seq2seq
Seq2seq is a classical model for structured learning, its input and output are both sequence
The vanilla seq2seq model is described in a NIPS '14 paper Sequence to Sequence Learning with Neural Networks, the encoder and the decoder are seperated
The seq2seq model in this repository is constructed with 2 LSTMs, similar to the one described in an ICCV '15 paper Sequence to Sequence -- Video to Text, the encoder and the decoder share same weights
- RL
After training chatbot with enough epochs, I use a RL technique called policy gradient to further improve the chatbot
By doing this, the chatbot can generate more interesting response with regard to the reward function
My reward function is similar to the one described in an EMNLP '16 paper Deep Reinforcement Learning for Dialogue Generation
Let me show you some chatbot's results:
- Result 1 (A & B are both a trained chatbot)
A: Can you do me a favor?
B: Anything.
A: I want you to bend a father.
B: Of course.
A: I'm so sorry I didn't mean to be rude but.
B: Don't get it at least it's about as much as I can.
- Result 2 (A & B are both a trained chatbot)
A: I didn't do it! I haven't got a gun!
B: You liar. You still want to get your money back.
A: Something wrong with the truth.
B: I don't want any money just a little I can't take just out of this.
A: What's wrong.
B: Listen to me. I'm scared now.
A: Ah you'll need all the money I need your help.
- Result 3 (A & B are both a trained chatbot)
A: I forgot to get the Coca-Cola.
B: I got something stuck in the head.
A: It's all right I guess it's pretty common in the lot of shit.
B: I mean we've been all this together since the kid.
A: All the more reason.
B: It's not him it's his fault and he's blind because of god.
A: And is a strange story.
you can find more results from chatbot in this directory
neural networks' configs of those results are described in the filename
Are you amazed at the result of the chatbot? :D
You can generate those amazing dialogs by your own!
Just follow the instructions below:
pip install -r requirements.txt./script/download.sh(Mention that I use -nc parameter in script/download.sh, it will omit downloading if the file exists
./script/simulate.sh <PATH TO MODEL> <SIMULATE TYPE> <INPUT FILE> <OUTPUT FILE>- <PATH TO MODEL>
to generate seq2seq dialog, type "model/Seq2Seq/model-77"
to generate RL dialog, type "model/RL/model-56-3000"
- <SIMULATE TYPE>
can be 1 or 2
the number represents # of former sentence(s) that chatbot considers
if you choose 1, chatbot only considers last sentence
if you choose 2, chatbot will consider last two sentences (one from user, and one from chatbot itself)
- <INPUT FILE>
Take a look at result/sample_input_new.txt
This is the input format of the chatbot, each line is the begin sentence of a dialog.
You can just use the example file for convenience.
- <OUTPUT FILE>
the output file, type any filename you want
If you want chatbot to generate only a single response for each question
Follow the instructions below:
pip install -r requirements.txt./script/download.sh(Mention that I use -nc parameter in script/download.sh, it will omit downloading if the file exists. So make sure there's no break during the download)
./script/run.sh <TYPE> <INPUT FILE> <OUTPUT FILE>- <TYPE>
to generate seq2seq response, type "S2S"
to generate reinforcement learning response, type "RL"
- <INPUT FILE>
Take a look at result/sample_input_new.txt
This is the input format of the chatbot, each line is the begin sentence of a dialog.
You can just use the example file for convenience.
- <OUTPUT FILE>
the output file, type any filename you want
I trained my chatbot with python2.7.
If you want to train the chatbot from scratch
You can follow those instructions below:
Take a look at python/config.py, all configs for training is described here.
You can change some training hyper-parameters, or just keep the original ones.
I use Cornell Movie-Dialogs Corpus
You need to download it, unzip it, and move all *.txt files into data/ directory
Then download some libraries with pip:
pip install -r requirements.txt./script/parse.sh./script/train.shLet's show some results of seq2seq model :)
./script/test.sh <PATH TO MODEL> <INPUT FILE> <OUTPUT FILE>And show some dialog results from seq2seq model!
./script/simulate.sh <PATH TO MODEL> <SIMULATE TYPE> <INPUT FILE> <OUTPUT FILE>- <SIMULATE TYPE>
can be 1 or 2
the number represents # of former sentence(s) that chatbot considers
if you choose 1, chatbot will only considers user's utterance
if you choose 2, chatbot will considers user's utterance and chatbot's last utterance
you need to change the training_type parameter in python/config.py
'normal' for seq2seq training, 'pg' for policy gradient
you need to first train with 'normal' for some epochs till stable (at least 30 epoches is highly recommended)
then change the method to 'pg' to optimize the reward function
./script/train_RL.shWhen training with policy gradient (pg)
you may need a reversed model
the reversed model is also trained by cornell movie-dialogs dataset, but with source and target reversed.
you can download pre-trained reversed model by
./script/download_reversed.shor you can train it by your-self
you don't need to change any setting about reversed model if you use pre-trained reversed model
Let's generate some results of RL model, and find the different from seq2seq model :)
./script/test_RL.sh <PATH TO MODEL> <INPUT FILE> <OUTPUT FILE>And show some dialog results from RL model!
./script/simulate.sh <PATH TO MODEL> <SIMULATE TYPE> <INPUT FILE> <OUTPUT FILE>- <SIMULATE TYPE>
can be 1 or 2
the number represents # of former sentence(s) that chatbot considers
if you choose 1, chatbot only considers last sentence
if you choose 2, chatbot will consider last two sentences (one from user, and one from chatbot itself)
- OS: CentOS Linux release 7.3.1611 (Core)
- CPU: Intel(R) Xeon(R) CPU E3-1230 v3 @ 3.30GHz
- GPU: GeForce GTX 1070 8GB
- Memory: 16GB DDR3
- Python3 (for data_parser.py) & Python2.7 (for others)
Po-Chih Huang / @pochih
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent