This is the official repository for our paper Learning to Generate Explanation from e-Hospital Services for Medical Suggestion, COLING 2022.
You can download our dataset from here.
The folder structure for the dataset is shown below, we saperate suggestion (recmd/) and explanation (expln/) in the testset of R2 and R3 for convinience to perform evaluation.
data
├── train
│ ├── R1_train.csv
│ ├── R2_train.csv
│ └── R3_train.csv
│
└── test
├── R1_test.csv
├── R2_test.csv
├── R3_test.csv
│
├── recmd
│ ├── R2_test_recmd.csv
│ └── R3_test_recmd.csv
│
└── expln
├── R2_test_expln.csv
└── R2_test_expln.csv
- Note
The dataset is originally collected from here.
- Install required packages listed in
requirements.txt. - Install
tw_rouge(for calculating Chinese Rouge score) by running the following command
pip install -e tw_rouge
[reference](https://github.com/cccntu/tw_rouge]
See scripts/ for our code.
All training and evaluation can be done by setting the correct arguments in train.sh and run
bash train.sh
-
python train.py \ --model_name_or_path google/mt5-base \ --train_data_path train/R1_train.csv \ --valid_data_path test/R1_test.csv \ --save_path results \ --do_train \ --do_eval \ --seed 42 \ --num_train_epochs 3 \ --per_device_train_batch_size 8 \ --per_device_eval_batch_size 32 \ --learning_rate 1e-4 \ --warmup_ratio 0.15 \ --weight_decay 0.01 \ --save_strategy epoch \ --logging_strategy steps \ --logging_steps 500 \ --evaluation_strategy epoch \ --metric_for_best_model rouge-1 \ --save_total_limit 1 \ --load_best_model_at_end \ --predict_with_generate True \ --generation_num_beams 2 \ -
python train.py \ --model_name_or_path google/mt5-base \ --train_data_path train/R2_train.csv \ --valid_data_path test/R2_test.csv \ --save_path results \ --do_train \ --do_eval \ --seed 42 \ --num_train_epochs 3 \ --per_device_train_batch_size 8 \ --per_device_eval_batch_size 32 \ --learning_rate 1e-4 \ --warmup_ratio 0.15 \ --weight_decay 0.01 \ --save_strategy epoch \ --logging_strategy steps \ --logging_steps 500 \ --evaluation_strategy epoch \ --metric_for_best_model rouge-1 \ --save_total_limit 1 \ --load_best_model_at_end \ --predict_with_generate True \ --generation_num_beams 2 \ -
python train.py \ --model_name_or_path google/mt5-base \ --train_data_path train/R3_train.csv \ --valid_data_path test/R3_test.csv \ --save_path results \ --do_train \ --do_eval \ --seed 42 \ --num_train_epochs 5 \ --per_device_train_batch_size 8 \ --per_device_eval_batch_size 32 \ --learning_rate 1e-4 \ --warmup_ratio 0.15 \ --weight_decay 0.01 \ --save_strategy epoch \ --logging_strategy steps \ --logging_steps 500 \ --evaluation_strategy epoch \ --metric_for_best_model rouge-1 \ --save_total_limit 1 \ --load_best_model_at_end \ --predict_with_generate True \ --generation_num_beams 2 \
Our arguments are directly compatible with the huggingface Seq2SeqTrainingArguments, you can adjust according to your own need. E.g., (un)set --do_train or do_eval if you only want to train or evaluate trained model.
If you want to evaluate only the suggestion/explanation for MTL mT5 or DMTL mT5, set --valid_data_path to the cooresponding testset data path as mentioned in the above Dataset section.