HS2DG-DST

This paper/code introduces the Multi-Domain Dialogue State Tracking via Dual Dynamic Graph with Hierarchical Slot Selector(HS2DG-DST)

The overview of HS2DG-DST:

Setting

You can use 0ysDST.yaml to import environment setting I used. Or get setting with code below.

# conda activate 0ysDST # python=3.8

conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
pip install transformers[sentencepiece]==4.24.0
pip install setproctitle

The pre-trained model we used is downloaded from huggingface

• Download the albert-base-v2 pre-training model, included config.json pytorch_model.bin spiece.model

• Put config.json pytorch_model.bin spiece.model into the pretrained_models/albert_large folder

Dataset

The schema/ontology is already in data folder. And you can download the dataset from https://github.com/budzianowski/multiwoz

preprocessed version

Thanks to guo, I used their preprocessed dataset.

• MultiWOZ 2.1: Google Drive
• MultiWOZ 2.2: Google Drive

Train & Evaluation

Before training, make sure you have prepared all input files(data/schema.json, data/train_dials.json, data/dev_dials.json, data/dev_dials.json) and pretrained models(pretrained_models/).

The train and evaluation code below.

$ conda activate 0ysDST

$ nohup ./train_model.sh > out/train_.out &
$ nohup ./eval_model.sh > out/eval_.out &

All model checkpoints will be saved to ./saved_models/.