Graph Evolving Meta-Learning for Low-resource Medical Dialogue Generation

Code to be further cleaned...

This repo contains the code of the following paper:

Shuai Lin, Pan Zhou, Xiaodan Liang, Jianheng Tang, Ruihui Zhao, Ziliang Chen, Liang Lin.
AAAI 2021

Prerequisites

Allennlp (0.9.1-unreleased)
pytorch == 1.4.0
Others should be found in ./allennlp/requirements.txt

[Note]: You need to install allennlp with the editable mode, i.e.,

cd ./allennlp
pip install --editable .
cd ..

since we have modified this toolkit (including added the metatrainer.py in the directory ./allennlp/training and so on).

Datasets

Please download both datasets from the google drive as follows:

wget https://drive.google.com/file/d/1KZ0CrIVZhSLxlZ-V5pnksvgH1xlyd54F/view?usp=sharing
tar zxvf cy.tar.gz
wget https://drive.google.com/file/d/1sZzb3Nzm_Z37lNCfgusJscFuiyhUON5j/view?usp=sharing
tar zxvf fd.tar.gz

CMDD: The directory fd/dis_pk_dir, which includes raw_data, meta_train and meta_test. (The number of the file name represents the ID of a disease.) You can also obtain it at the link
MDG-Chunyu: The directory cy/dis_pk_dir, which also includes the raw_data, meta_train and meta_test. The ID of diseases and symptoms are recorded in the user_dict.txt. The disease IDs are as follows:

{
  '胃炎': 2,
  '普通感冒': 13,
  '肺炎': 73,
  '便秘': 6,
  '胃肠功能紊乱': 42,
  '肠炎': 9,
  '肠易激综合征': 40,
  '食管炎': 27,
  '胃溃疡': 30,
  '阑尾炎': 35,
  '胆囊炎': 33,
  '胰腺炎': 48,
  '肠梗阻': 52,
  '痔疮': 18,
  '肝硬化': 46,
}

Quick Start

Most of the running commands are written in the script run.sh, which follows the offical train/fine-tune/evaluate way of the allennlp. Take the following one as an example:

[1]. Training:

CUDA_VISIBLE_DEVICES=1 allennlp train -s $save_directory$ \
  $config_file(.json)$ \
  --include-package $model_file$

[2]. Fine-tuning:

CUDA_VISIBLE_DEVICES=1 allennlp fine-tune -m $old save_directory$ \
  -c $config_file(.json)$ \
  --include-package $model_file$
  -s $new save_directory$

[3]. Testing:

CUDA_VISIBLE_DEVICES=3 allennlp evaluate  $new save_directory$ \
  $test_data$ \
  --include-package $model_file$ \
  --output-file $output_directory$

ha-lins/GEML-MDG

Graph Evolving Meta-Learning for Low-resource Medical Dialogue Generation

Prerequisites

Datasets

Quick Start