dialog-rl's Introduction

Domain Transfer in Dialogue Systems without Turn-Level Supervision

This is the code repository for our paper Domain Transfer in Dialogue Systems without Turn-Level Supervision.

Run it

Install dependencies

pip install -r requirements.txt

Run training script

python run.py

This script gives you plenty of options for arguments controlling training, and it's also where you specifiy what data you'll work on. Please run python run.py -h for an overview of options.

Evaluating

If you just want to evaluate a trained model and produce predictions, you do:

python run.py --test

Cite

If you use this code, please cite the paper. Bibtex:

@article{bingel2019domain,
  title={Domain Transfer in Dialogue Systems without Turn-Level Supervision},
  author={Bingel, Joachim and Petr\'en Bach Hansen, Victor and Gonzalez, Ana Valeria and Budzianowski, Pawe{\l}  and Augenstein, Isabelle and S{\o}gaard, Anders},
  journal={arXiv preprint arXiv:1909.07101},
  year={2019}
}

dialog-rl's People

Contributors

Stargazers

Watchers

dialog-rl's Issues

Cant find "emb.json" files

I am trying to run the code like this :

! python run.py --epochs 10 --lr 0.005 --reinforce --gpu 0 --train_domains ['taxi'] --finetune_domains ['train'] --eval_domains ['train'] --path "/content/dialog-rl/Preprocessing/ann/" --outfile "/content/outfile"

And I get this error:

FileNotFoundError: [Errno 2] No such file or directory: '/content/dialog-rl/Preprocessing/ann/emb.json'

there is no data folder but I thought the files are in Preprocessing folder hence the path. Could someone tell me what I am doing wrong? I tried running the preprocess_data.py file but I run into a bunch of errors there too.

Implement pretrain/finetune strategy

Make sure datasets are read correctly and see whether non-identical sets of vocab/embeddigns/etc aren't problematic

Allow for loading pretrained model from disk for RL finetuning experiments

None option for slots

Delexicalize postcodes, reservation numbers, phone numbers etc.

e.g. replace with dummy placeholder

Modify load_dataset to allow for domain filtering

We talked about a setup that pretrains on domains D_1 ... D_{n-1} and then finetunes on domain D_n.

This requires loading the datasets separately for those domains. I've rewritten the header of load_dataset() to reflect this. The idea is that domains is a list of strings with domain names -- but if it's just a single string 'all', everything should pass. The function also devises a strict flag that only allows dialogs to pass that are one domain only.