Generative Chit-Chat

Data preparation

The model was trained on Ответы Mail.ru dataset (1000K examples) tokenized with pre-trained BPE-tokenizer youtokentome.

Model

• Encoder: LSTM
• Decoder: LSTM with Attention Mechanism

To run training:

python train.py

Evaluation

For decoding answers greedy-search decoder is used. To see results:

python test.py

Examples

See examples of model answers in test.tsv file.