hanzhou032/xqa-dst

XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking (Zhou et al.; EACL 2023 Findings)

Code for XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking.

Link to paper: XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking

Authors: Han Zhou, Ignacio Iacobacci, Pasquale Minervini

In Findings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2023.

Dependencies

Install pytorch, transformers, tensorboardX, and Googletrans

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch-nightly -c nvidia
pip install transformers==2.9.1
conda install tensorboardX
pip install googletrans==3.1.0a0

XQA-DST Experiments in the paper

Zero-shot Domain Adaptation Experiments

1. First define the excluded domain in the file examples/example.train.xlmr.multiwoz.adaptation
2. Execute the following script that will do the leave-1-out domain adaptation experiment for you.

sh examples/example.train.xlmr.multiwoz.adaptation 

Zero-shot Cross-Lingual Experiments

1. First train the XLM-R on WOZ 2.0 English datasets. We start from the checkpoint of the XLM-R trained on SQuAD 2.0

sh examples/example.train.xlmr.squad

2. Define the best learned checkpoint in 'model_name_or_path'. Transfer it to GE and IT. Notice the Googletrans process is time-consuming.

sh examples/example.test.cross.lingual.it
sh examples/example.test.cross.lingual.de

Supervised DST Experiments

1. We first train the independent domain classifier using files ending with domain_classifier_multiple

sh examples/example.train.xlmr.multiwoz.classifier

2. Then train the main model independently.

sh examples/example.train.xlmr.multiwoz

3. Then load them together within the arguments of the file example.test.xlmr.combine, which should generate evaluation results.

sh examples/example.test.xlmr.combine

Datasets

Supported datasets are:

• WOZ 2.0
• MultiWOZ 2.1

The prepared and processed datasets for this work is 'preprocessed_data.zip'. Please unzip it to the data/ folder for your convenience.

Partial codes are modified from TripPy: A Triple Copy Strategy for Value Independent Neural Dialog State Tracking

Backlog:

1. Add support for new multi-lingual ToD datasets.
2. Replace the framework to the newest version of Huggingface transformers.

Citation

If you find our work to be useful, please cite:

@article{zhou2022xqa,
  title={XQA-DST: Multi-Domain and Multi-Lingual Dialogue State Tracking},
  author={Zhou, Han and Iacobacci, Ignacio and Minervini, Pasquale},
  journal={arXiv preprint arXiv:2204.05895},
  year={2022}
}