uralik / beamdream

Importance of a Search Strategy in Neural Dialogue Modelling

https://uralik.github.io/beamdream/

This repo provides code, trained models, run scripts and human evaluation transcripts for our work on different search strategies for neural dialogue models.

Code

All the training and human evaluation was done using ParlAI framework. ParlAI is being actively developed and we do not confirm our code to be working with master branch code.

It should be working using ParlAI commit 5899c07934836d8757ebfc8d98973bdef2c56c74

We plan to include scorer functionality together with scorer in main ParlAI repo in the near future.

To use seq2seq_on_steroids agent please do the following:

git clone [email protected]:facebookresearch/ParlAI.git
git clone [email protected]:uralik/beamybeam.git
cp -r beamybeam/parlai_external ParlAI/
cd ParlAI; python setup.py develop

After that you should be able to import SteroidSeq2seqAgent using this command:

from parlai_external.agents.seq2seq_on_steroids.seq2seq_on_steroids import SteroidSeq2seqAgent
from parlai_external.agents.seq2seq_on_steroids.modules import SteroidSeq2seq

Trained models

This is the model used for all experiments in the paper. Corresponding .opt file provides all hyperparameters which were used during the training.

seq2seq model

Archive contains typical set of files needed in ParlAI to do any kinds of further tasks. Please see ParlAI docs for further details.

To make a quick check to verify your model is running you can use eval script after paths adjustments.

Citation

Please use the following bib if you wish to cite our work:

@misc{kulikov2018importance,
    title={Importance of Search and Evaluation Strategies in Neural Dialogue Modeling},
    author={Ilia Kulikov and Alexander H. Miller and Kyunghyun Cho and Jason Weston},
    year={2018},
    eprint={1811.00907},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

beamdream's Issues

similarity threshold used for paper

I was wondering what value is being used for the $$\epsilon$$ in equation 3 of the paper.
I see that a variable min_hamming_dist gets passed into advance in beam_on_steroids.py, but it doesn't look like it actually gets used.
Was $$\epsilon=1$$ used for all the experiments in the paper? In this case, wouldn't this mean no relaxation is happening?

Model link is broken

Hi
First of thanks for great work
I was trying to test your model but the model file are not available. Is it possible to fix the links?

Thanks