DEPRECATED

This project was built in 2016 (the old tensorflow era), and some design choices no longer make sense under the new tensorflow. You are likely to find better seq2seq tutorial/implementation somewhere else.

seq2seq_chatbot

An implementation of Seq2seq chatbot in tensorflow.

Features

• dynamic rnn with smart loader (padding free)
• beam search on prediction (fast approximation on global optimum)
• signal indicator for decoder (partial control on decoder)

A technical report: https://docs.google.com/gview?url=http://sudongqi.com/Documents/2016_02.pdf&embedded=true

Python 2.7 dependency

• tensorflow 1.8
• numpy
• json

Instruction

• run "python train.py" and wait (5 minutes on GTX 1080 Ti with cuda 9.0 and cudnn 7.0) until training is completed
• run "python test.py" to enter the interactive session with the chatbot

Try your own data

it's possible to run it on your own data, but you need to generate at least 2 files with the same format like those in bbt_data.

• text.txt this is the training data contatining the pair in number token format
• dict.json this is the dicitonary to translate from number token to English word token in test time
• actors.json (optional) this is for signal indication in test time
• summary.json (optional) this file contain the length info for selecting the right bucket options for training

OpenSubtitles data

If you want to train on openSubtitles (english 2016) dataset, this project provide a data processing script (data_processing/openSub_data_generator.py) for openSubtitles. Get OpenSubtitles data from here: http://opus.lingfil.uu.se/OpenSubtitles2016.php