Doctor-Recommendation

The official implementation of ACL 2022 paper "Doctor Recommendation in Online Health Forums via Expertise Learning".

Dataset

Our dataset (avaliable at dataset directory) was collected from Chunyu Yisheng(春雨医生). Our dataset is collected by a crawler within the constraints of the forum. Apart from the personal information de-identified by the forum officially, to prevent privacy leaks, we manually reviewed the collected data and deleted sensitive messages.

Data format

• embed.csv, train.csv, valid.csv, test.csv
  • train.csv, valid.csv, test.csv are the training, validation, testing splits of our dataset respectively. embed.csv is the combination of these three csv files i.e., total data.
  • they have the same colums:
    • "dr_id": doctor ID
    • "dialog_id": ID of both dialogues and queries.
    • "q": query content
    • "parsed_dialog": parsed dialogues (for a dialogue d, we convert it into a token sequence via linking turns in chronological order.)
• dialogues.json: dialogues with raw format
• dr_profile.jsonl: doctor informations (we use "goodat" of each doctor as profile.)

Data statistics

# of dialogues	119,128
# of doctors	359
# of departments	14
# of tokens in vocabulary	8,715
Avg. # of dialogues per doctor	331.83
Avg. # of doctors per department	25.64
Avg. # of tokens in a query	89.97
Avg. # of tokens in a dialogue	534.28
Avg. # of tokens in a profile	87.53

Model

Codes

Dependencies

1. Make sure having python libarary virtualenv installed, or install with pip install virtualenv.
2. Initialize a new virtualenv and install all dependencies.

   python -m venv env # create virtual environment 
   source env/bin/activate # activate virtual environment. 
   pip install -r requirements.txt # install all dependencies

Self-Learning

This self-learning task is to predict whether a profile and a dialogue come from the same doctor, where random profile-doctor pairs are adopted as the negative samples. We first fine-tuned mc_bert_base (a pre-trained Chinese Biomedical BERT) via self-learning.

The dataset for self-learning is avaliable at self-learning/dataset. To run both training and evaluation of self-learing task, turn to the self-learning directory, run:

python self_learning.py -seed 2021 -epoch_num 20 -batch_size 50 -accumulation_steps 5

Checkpoints will be stored in self_learning/checkpoints directory. We choose our best self-learning checkpoint and move it into 'sl_best_model' that will be used later.

Bert Embedding

We employ a pre-trained MC-BERT (fine-tuned via self-learning) to encode profile, dialogues, queries and obtain their rudimentary embeddings: dialog_embeddings.json, profile_embeddings.json and q_embeddings.json.

# load self-learning finetuned model from sl_best_model
# output embeddings path: bert_embeddings
python embed.py -load_sl_model 1 
# load mc_bert_base model (i.e., without finetuning) from mc_bert_base
# output embeddings path: bert_embeddings_wo_sl
python embed.py -load_sl_model 0  

Multi-head Attention (MUL-ATT) and Recommendation Prediction.

MUL-ATT: With embeddings of doctor profiles, dialogues and queries, it then employs profile-aware multi-head attention over dialogues to explore doctor expertise and works with the query encoder (to capture patient needs) to pair doctors with queries.

Recommendation Prediction: Given a pair of doctor $D$ and query $q$, the embedding results of doctor encoder $e_D$ and query encoder $e_q$ are coupled in the prediction layer for recommendation. We adopt a MLP architecture to measure the matching score $s$ of the $D-q$ pair, which indicates the likelihood of doctor $D$ able to provide a suitable answer to query $q$

We provide three bashs scripts train.sh, test.sh and eval.sh to run the training, prediction, and evaluation of three MUL-ATT models:

• MUL-ATT (W/O SL): multi-head attention without this self-learning step
• MUL-ATT (W/O D): encode profiles only with a multi-head self-attention
• MUL-ATT (W/O P): with dialogues only
• MUL-ATT (FULL): the full model

train.sh and predict.sh will call python train.py and python predict.py respectively for training and prediction. The experiment settings in train.sh and predict.sh are corresponding to settings stated in config.py.

Note that all experiments will be run in parallel and select a single available GPU sequentially. You can change the number of total gpus (i.e. n_gpu) in train.sh and predict.sh. To monitor experiments, you may view the corresponding generated log file.

For the evaluation (i.e., eval.sh), we use RankLib to evaluate the predictions with information retrieval metrics: precision@N (P@N), mean average precision (MAP), and ERR@N . N is set to 1 for P@N and 5 for ERR@N.

Citation

@inproceedings{lu-etal-2022-doctor,
    title = "Doctor Recommendation in Online Health Forums via Expertise Learning",
    author = "Lu, Xiaoxin  and
      Zhang, Yubo  and
      Li, Jing  and
      Zong, Shi",
    booktitle = "Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = may,
    year = "2022",
    address = "Dublin, Ireland",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.acl-long.79",
    pages = "1111--1123",
}