VRBot: Semi-Supervised Variational Reasoning for Medical Dialogue Generation

This paper has been accepted by SIGIR 2021 as a full paper, the code is based on PyTorch, and the paper is here.

Model

• probabilistic graphical model

• model overview

Dataset

You can access the dataset via the following two links:

• Baidu Link: https://pan.baidu.com/s/1uEZhLC4gxq7Fn1u8VrGgPw Password: dmpm
• Google Link: https://drive.google.com/drive/folders/1i-qiwVgOHS9Cs_7YSNdUCWwviP2HOgqI?usp=sharing

Requirements

• pytorch 1.4.0
• tqdm 4.31.1
• numpy 1.16.2
• pandas 0.24.0
• logging 0.5.1.2
• argparse 1.1
• json 2.0.9
• nlgeval

Notations

• The data should be loaded with several processes in parallel, please make sure your CPU with over 4 cores.
• We conduct our experiments on RTX 2080ti(11GB) and Titan RTX(24GB), please make sure your GPU with over 10 GB of memory.

Experiments

• train

python main.py --task [kamed,meddialog,meddg] --super_rate 0.5 --model vrbot --train_batch_size 16 --test_batch_size 32 --device 0 --worker_num 5

• test

python main.py --task [kamed,meddialog,meddg] --test --model vrbot --test_batch_size 32 --device 0 --ckpt ckpt_filename --worker_num 5

• evaluation (compute the metrics and t-test):

python eval_main.py --eval_filename "your_output_filename" --vocab_filename "your_vocab_filename" --alias2scientific_filename "your_alias_filename"

python eval_main.py --eval_filename "your_output_filename" --eval_filename2 "your_second_output_filename" --vocab_filename "your_vocab_filename" --alias2scientific_filename "your_alias_filename"

please download the word embedding via link: https://pan.baidu.com/s/1A-EzSJUdSu6g8Zy8BCm-Cw password: qw9i .

Reference

If you use any code or data in this repo in your work, please cite our paper.

@inproceedings{li2021semi,
title={Semi-Supervised Variational Reasoning for Medical Dialogue Generation},
author={Li, Dongdong and Ren, Zhaochun and Ren, Pengjie and Chen, Zhumin and Fan, Miao and Ma, Jun and de Rijke, Maarten},
year = {2021},
booktitle = {SIGIR},
pages = {},
}