lixiangnlp / seq2seq-1

sequence to sequence neural network model for dialogue system

seq2seq

Seq2seq is a Python library for modeling dialogue/conversational model with neural network.

This software implements seqence to sequence (seq2seq) neural network models with the aim to create dialogue systems.

Install

This software depends on chainer and gensim. You must install them first.

$ pip3 install chainer==1.6.1
$ pip3 install gensim==0.12.4

It is also recommended to install cuda to use GPU. It is optional, but using GPU improves performance.

$ pacman -S cuda  # for ArchLinux

Usage

You must learn a model to use seq2seq first with you own corpus. It is easy to use your own corpus in this module. The following description uses test corpus included in this software to show you how to apply learning scripts to the corpus.

corpus_test directory has a test corpus conv.txt. corpus should be the folllowing format.

original_ sentence<TAB>reply_sentence

A sentence in the second column is reply to a sentence in the first column. These sentences should be separated by TAB.

Each sentence should be divided into tokens such as words or characters. You can utilize corpus_test/Makefile to split sentences into words and characters.

$ cd corpus_test
$ make char

Makefile splits conv.txt to sent.char.txt and conv.char.txt. sent.char.txt has all texts in conv.txt splitted by characters. conv.char.txt has all conversations in conv.txt splitted by characters.

Makefile can also split conv.txt to sent.word.txt and conv.word.txt by using make word. When you use this, you should specify TOKENIZER in Makefile first. The default is mecab -Owakati. sent.word.txt has all texts in conv.txt splitted by words. conv.word.txt has all conversations in conv.txt splitted by words.

This section uses conv.sent.txt and sent.sent.txt to descrive usage, but you can also use conv.word.txt and sent.word.txt instead.

After preparing a corpus, you can learn your model which predicts a reply sentence from a input sentence based on the corpus. There are a configuration file test.ini which has parameters to learn a model from the test corpus.

Use train.py to learn your model.

$ python train.py test.ini -tlstm

Specify -g0 if you use GPU.

$ python train.py test.ini -tlstm

After finishing training, use test.py for talking with the model.

$ python test.py test.ini -tlstm <./corpus_test/sent.char.txt

Use -g option to use GPU.

# enable GPU to use -g option
$ python train.py test.ini -g0 -tlstm
$ python test.py test.ini -g0 -tlstm

There is Makefile for convenient.

# train the model
$ make train
# test the model
# make test

Model description

• Word Embedding initializer: word2vec or random between [0, 1]
• Layers : embedding layer -> 2 hidden layers -> output layer -> Softmax
• Units/activation functions: LSTM or ReLU with dropout option
• Optimizer: ADAM with clipping