HyKnow

This is the code for paper HyKnow: End-to-End Task-Oriented Dialog Modeling with Hybrid Knowledge Management.

If you use any source code or dataset included in this repo in your work, please cite this paper:

@inproceedings{gao2021hyknow,
  title={HyKnow: End-to-End Task-Oriented Dialog Modeling with Hybrid Knowledge Management},
  author={Gao, Silin and Takanobu, Ryuichi and Peng, Wei and Liu, Qun and Huang, Minlie},
  booktitle={ACL-IJCNLP: Findings},
  pages={1591--1602},
  year={2021}
}

Requirements

• Python 3.6
• PyTorch 1.2.0
• NLTK 3.4.5
• Spacy 2.2.2

We use some NLP tools in NLTK which can be installed through:

python -m nltk.downloader stopwords punkt wordnet

Dataset

1. Raw dataset: modified MultiWOZ 2.1

2. Our preprocessed dataset can be downloaded from this link, please unzip the file under the root directory and data is placed in data/.

3. Pretrained Glove embeddings can be downloaded from this link, please unzip the file and place it in data/.

Implementations of HyKnow

We build HyKnow in both single-decoder and multi-decoder belief state decoding implementations.

HyKnow with single-decoder belief state decoding implementation: HyKnow_Single/ HyKnow with multi-decoder belief state decoding implementation: HyKnow_Multiple/

Running Experiments

Before running, place the preprocessed dataset data/ into HyKnow_Single/ or HyKnow_Multiple/. Go to the experiment root:

cd HyKnow_Single

or

cd HyKnow_Multiple

Training

python train.py -mode train -dataset multiwoz -method bssmc -c spv_proportion=100 exp_no=your_exp_name

Testing

python train.py -mode test -dataset multiwoz -method bssmc -c eval_load_path=[experimental path]

Best Results

We release the best results obtained by the two implementations of our model. Please unzip the file from this link, and results are placed in results/