diversity-promoting-dialogue-system

This is a implementation with Pytorch and Python for diversity-promoting dialogue system proposed by Li et al. 2015

Requirement:

Python 3.6.4 (not limited to)
Jupyter Notebook
Cuda 8.0
Pytorch 0.4.0

Corpus:

The dataset folder contains a tweets dialogue file round 700k lines. You can also replace with own dataset, as long as the input-output sequence pairs are in the adjacent lines. For example,

message1
response1
message2
response2

Train:

Two jupyter notebooks correspond to MMI-antiLM model and MMI-bidi model. To run it, please download or clone this repository.

When getting started, set train_type variable as 'restart' in dataloading part (block No.5). During training, you can save dataset partition as well as network parameters for future use, by simply setting train_type as 'resume'.

Evaluation:

Beam search is built for decoding, and the main measurements are BLEU score and distinct value of unigrams and bigrams. In notebook, there are some records with different hyper-parameters I tried. One can also play with other combination.
Evaluation on baseline seq2seq model achieves BLEU 3.73 on test set, with 0.103 and 0.412 for distinct-1 and distinct-2.
For MMI-antiLM, evaluation on test set achieves BLEU 4.21, with 0.126 and 0.457 for distinct-1 and distinct-2.
For MMI-bidi, evaluation on test set achieves BLEU 4.20, with 0.111 and 0.408 for distinct-1 and distinct-2.

Result (sample):

Here are a few results sampled from test set.

MMI-antiLM -- gamma=0.2, lambda=0.2, threshold=2

MMI-bidi -- gamma=-0.2, lambda=0.5

TODO:

1. Find better teacher forcing ratio (current version use 100% ratio, i.e. completely follows reference policy)
2. Tune dropout value (current version's dropout=0)