DST-SC

This repository is the implementation of Dialogue State Tracking with Explicit Slot Connection Modeling.

Requirements

Install requirements:

pip install -r requirements.txt

Other preparations:

• Unzip dataset/multiwoz.zip
• Download character embedding and unzip charNgram.txt to the embedding folder.

Training

To train the model in the paper, run this command:

python run.py --dataset=2.0 --gpu=0,1,2 --batch_size=6 --gas=2

Evaluation

To evaluate the model, specify your saved checkpoint file in train.py first and run:

python run.py --dataset=2.0 --is_test=True

Model

Results

Our model achieves the following performance on MultiWOZ 2.0 and 2.1 dataset:

Model	MultiWOZ 2.0	MultiWOZ 2.1
DST-SC	52.24%	49.58%

Citation

@inproceedings{ouyang-etal-2020-dialogue,
    title = "Dialogue State Tracking with Explicit Slot Connection Modeling",
    author = "Ouyang, Yawen  and
      Chen, Moxin  and
      Dai, Xinyu  and
      Zhao, Yinggong  and
      Huang, Shujian  and
      Chen, Jiajun",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.5",
    doi = "10.18653/v1/2020.acl-main.5",
    pages = "34--40",
    abstract = "Recent proposed approaches have made promising progress in dialogue state tracking (DST). However, in multi-domain scenarios, ellipsis and reference are frequently adopted by users to express values that have been mentioned by slots from other domains. To handle these phenomena, we propose a Dialogue State Tracking with Slot Connections (DST-SC) model to explicitly consider slot correlations across different domains. Given a target slot, the slot connecting mechanism in DST-SC can infer its source slot and copy the source slot value directly, thus significantly reducing the difficulty of learning and reasoning. Experimental results verify the benefits of explicit slot connection modeling, and our model achieves state-of-the-art performance on MultiWOZ 2.0 and MultiWOZ 2.1 datasets.",
}