Skip to content
Branch: revert-10-mast…
Clone or download


the paper: Adversarial Learning for Neural Dialogue Generation

the paper translation in Chinese :


TensorFlow 0.12.0 Python 2.7

Introduction to Project: al_neural_dialogue


gen_data:   training data for gen model

disc_data:   training data for disc model

disc:       code about disc model

gen:         code about gen model

utils:       code about data operation and model config


gen_data include chitchat.train.answer, chitchat.train.query,, (total four files)

disc_data include,, 和 disc.train.answer, disc.train.query,disc.tran.gen (total six files)

formula of training data one sentence one row and splited with space, eg: i don ' t want to !



def main(_):


  # step_1 training gen model

# gen_pre_train()

# model test
# gen_test()

# step_2 gen training data for disc
# gen_disc()

# step_3 training disc model
# disc_pre_train()

# step_4 training al model
# al_train()

# model test
# gen_test() 


model introduction

1、disc model : hierarchical rnn (paper——Building end-to-end dialogue systems using generative hierarchical neural network models)

2、gen model : seq2seq model with attention (GRU cell)

3、method of reward : Monte Carlo Search

4、optimal:Policy Gradient

You can’t perform that action at this time.