Dialogue Knowledge Transfer Networks (DiKTNet)

This codebase is shared by the following papers: "Few-Shot Dialogue Generation Without Annotated Data: A Transfer Learning Approach" by Igor Shalyminov, Sungjin Lee, Arash Eshghi, and Oliver Lemon. [SigDial 2019 publication] [bibtex] [Poster]

"Data-Efficient Goal-Oriented Conversation with Dialogue Knowledge Transfer Networks" by Igor Shalyminov, Sungjin Lee, Arash Eshghi, and Oliver Lemon. [EMNLP 2019 publication] [bibtex] [Poster]

If you find it useful for your work, please cite the papers above.

The versions of the code corresponding to each publication can be found using the git release tags.

Pre-requisites

1. Download the Maluuba MetaLWOz (previously called BLIS) dataset (assuming you extract it into metalwoz-v1 folder)
2. Prepare environment and dependencies. Below are the steps for Conda:

  conda create -n diktnet python=3.6
  conda activate diktnet
  git submodule update --init
  pip install -r requirements.txt

3. Create a vocabulary for LAED training: python make_vocabulary.py <data-dir> <corpus client type> <vocab_file.json>

Given the data sources we have, it may be NeuralDialog-ZSDG/data/stanford/ZslStanfordCorpus, NeuralDialog-ZSDG/data/simdial/SimDialCorpus or metalwoz-v1/LAEDBlisCorpus. Vocabularies of a higher coverage can be produced by merging the primary ones.

How To

1. Training a LAED model - StED (skip-thought dialog-level)

python st_ed.py \
    metalwoz-v1 \
    LAEDBlisCorpus \
    vocabs/maluuba.json \
   --exlcude_domains ... \
   --y_size 10 

1.1 Training a LAED model - vanilla VAE (dialog-level)

python vae.py \
  metalwoz-v1 \
  LAEDBlisCorpus \
  vocabs/maluuba.json \
 --exlcude_domains ... \
 --y_size 1
 --k 100

2. Generating LAED features - StED (skip-thought dialog-level)

python generate_laed_features.py \
    laed_models/st_ed_maluuba_${TARGET_DOMAIN} \
    laed_features/st_ed_maluuba__smd_${TARGET_DOMAIN} \
    --model_name StED \
    --model_type dialog \
    --data_dir NeuralDialog-ZSDG/data/stanford \
    --corpus_client ZslStanfordCorpus \
    --data_loader SMDDialogSkipLoader \
    --vocab maluuba.json

3. Training a Few-Shot Dialog Generation model

python train_fsdg.py \
    LAZslStanfordCorpus \
    --data_dir NeuralDialog-ZSDG/data/stanford \
    --laed_z_folders laed_features/st_ed_3x5_maluuba__smd_${domain} \
    --black_domains $domain \
    --black_ratio 0.9 \
    --action_match False \
    --target_example_cnt 0 \
    --random_seed $rnd