English | 中文说明

CBLUE

AI (Artificial Intelligence) plays an indispensable role in the biomedical field, helping improve medical technology. For further accelerating AI research in the biomedical field, we present Chinese Biomedical Language Understanding Evaluation (CBLUE), including datasets collected from real-world biomedical scenarios, baseline models, and an online platform for model evaluation, comparison, and analysis.

CBLUE Benchmark

We evaluate the current 11 Chinese pre-trained models on the eight biomedical language understanding tasks and report the baselines of these tasks.

Baseline of tasks

We present the baseline models on the biomedical tasks and release corresponding codes for a quick start.

Requirements

python3 / pytorch 1.7 / transformers 4.5.1 / jieba / gensim / sklearn

Data preparation

Download dataset

The whole zip package includes the datasets of 8 biomedical NLU tasks (more detail in the following section). Every task includes the following files:

├── {Task}
|  └── {Task}_train.json
|  └── {Task}_test.json
|  └── {Task}_dev.json
|  └── example_gold.json
|  └── example_pred.json
|  └── README.md

Notice: a few tasks have additional files, e.g. it includes 'category.xlsx' file in the CHIP-CTC task.

You can download Chinese pre-trained models according to your need (download URLs are provided above). With Huggingface-Transformers , the models above could be easily accessed and loaded.

The reference directory:

├── CBLUE         
|  └── baselines
|     └── run_classifier.py
|     └── ...
|  └── examples
|     └── run_qqr.sh
|     └── ...
|  └── cblue
|  └── CBLUEDatasets
|     └── KUAKE-QQR
|     └── ...
|  └── data
|     └── output
|     └── model_data
|        └── bert-base
|        └── ...
|     └── result_output
|        └── KUAKE-QQR_test.json
|        └── ...

Running examples

The shell files of training and evaluation for every task are provided in examples/ , and could directly run.

Also, you can utilize the running codes in baselines/ , and write your shell files according to your need:

• baselines/run_classifer.py: support {sts, qqr, qtr, qic, ctc, ee} tasks;
• baselines/run_cdn.py: support {cdn} task;
• baselines/run_ie.py: support {ie} task.

Training models

Running shell files: bash examples/run_{task}.sh, and the contents of shell files are as follow:

DATA_DIR="CBLUEDatasets"

TASK_NAME="qqr"
MODEL_TYPE="bert"
MODEL_DIR="data/model_data"
MODEL_NAME="chinese-bert-wwm"
OUTPUT_DIR="data/output"
RESULT_OUTPUT_DIR="data/result_output"

MAX_LENGTH=128

python baselines/run_classifier.py \
    --data_dir=${DATA_DIR} \
    --model_type=${MODEL_TYPE} \
    --model_dir=${MODEL_DIR} \
    --model_name=${MODEL_NAME} \
    --task_name=${TASK_NAME} \
    --output_dir=${OUTPUT_DIR} \
    --result_output_dir=${RESULT_OUTPUT_DIR} \
    --do_train \
    --max_length=${MAX_LENGTH} \
    --train_batch_size=16 \
    --eval_batch_size=16 \
    --learning_rate=3e-5 \
    --epochs=3 \
    --warmup_proportion=0.1 \
    --earlystop_patience=3 \
    --logging_steps=250 \
    --save_steps=250 \
    --seed=2021

Notice: the best checkpoint is saved in OUTPUT_DIR/MODEL_NAME/.

• MODEL_TYPE: support {bert, roberta, albert, zen} model types;
• MODEL_NAME: support {bert-base, bert-wwm-ext, albert-tiny, albert-xxlarge, zen, pcl-medbert, roberta-large, roberta-wwm-ext-base, roberta-wwm-ext-large, macbert-base, macbert-large} Chinese pre-trained models.

The MODEL_TYPE-MODEL_NAME mappings are listed below.

Inference & generation of results

Running shell files: base examples/run_{task}.sh predict, and the contents of shell files are as follows:

DATA_DIR="CBLUEDatasets"

TASK_NAME="qqr"
MODEL_TYPE="bert"
MODEL_DIR="data/model_data"
MODEL_NAME="chinese-bert-wwm"
OUTPUT_DIR="data/output"
RESULT_OUTPUT_DIR="data/result_output"

MAX_LENGTH=128

python baselines/run_classifier.py \
    --data_dir=${DATA_DIR} \
    --model_type=${MODEL_TYPE} \
    --model_name=${MODEL_NAME} \
    --model_dir=${MODEL_DIR} \
    --task_name=${TASK_NAME} \
    --output_dir=${OUTPUT_DIR} \
    --result_output_dir=${RESULT_OUTPUT_DIR} \
    --do_predict \
    --max_length=${MAX_LENGTH} \
    --eval_batch_size=16 \
    --seed=2021

Notice: the result of prediction {TASK_NAME}_test.json will be generated in RESULT_OUTPUT_DIR .

Check format

Before you submit the predicted test files, you could check the format of test files using format_checker and avoid the invalid evalution score induced by the format errors.

• Step1: Copy the original test file(without answer) {taskname}_test.[json|jsonl|tsv] to this directory format_checker, and rename as {taskname}_test_raw.[json|jsonl|tsv].

  # take the CMeEE task for example:
  cp ${path_to_CMeEE}/CMeEE_test.json ${current_dir}/CMeEE_test_raw.json 

• Step2: Execute the following format_checker script using the raw test file (from Step1) and your prediction file:

  
  python3 format_checker_${taskname}.py {taskname}_test_raw.[json|jsonl|tsv] {taskname}_test.[json|jsonl|tsv] 

take the CMeEE task for example:

python3 format_checker_CMeEE.py CMeEE_test_raw.json CMeEE_test.json

#### What is special? 
##### IMCS-NER & IMCS-V2-NER tasks:
* Step1: Copy both the original test file(without answer) IMCS-NER_test.json(IMCS-V2-NER_test.json) and the IMCS_test.json(IMCS-V2_test.json) to this directory, and rename as IMCS-NER_test_raw.json(IMCS-V2-NER_test_raw.json)
```shell
# for IMCS-NER task:
cp ${path_to_IMCS-NER}/IMCS-NER_test.json ${current_dir}/IMCS-NER_test_raw.json 
cp ${path_to_IMCS-NER}/IMCS_test.json ${current_dir}
# for IMCS-V2-NER task:
cp ${path_to_IMCS-V2-NER}/IMCS-V2-NER_test.json ${current_dir}/IMCS-V2-NER_test_raw.json 
cp ${path_to_IMCS-V2-NER}/IMCS-V2_test.json ${current_dir}

• Step2: Execute the following format_checker script using the raw test file (from Step1) and your prediction file:

  # for IMCS-NER task:
  python3 format_checker_IMCS_V1_NER.py  IMCS-NER_test_raw.json IMCS-NER_test.json IMCS_test.json
  # for IMCS-V2-NER task:
  python3 format_checker_IMCS_V2_NER.py  IMCS-V2-NER_test_raw.json IMCS-V2-NER_test.json IMCS-V2_test.json

  IMCS-SR & IMCS-V2-SR, MedDG tasks

  If you want to implement the optional check login in the check_format function, which is commented in the master branch. You need also copy the normalized dictionary files to the current dir.

• MedDG: the dictionary file is entity_list.txt
• IMCS-SR: the dictionary file is symptom_norm.csv
• IMCS-V2-SR: the dictionary file is mappings.json

Submit results

Compressing RESULT_OUTPUT_DIR as .zip file and submitting the file, you will get the score of evaluation on these biomedical NLU tasks, and your ranking!

Submit your results!

Introduction of tasks

For promoting the development and the application of language model in the biomedical field, we collect data from real-world biomedical scenarios and release the eight biomedical NLU (natural language understanding) tasks, including information extraction from the medical text (named entity recognition, relation extraction), normalization of the medical term, medical text classification, medical sentence similarity estimation and medical QA.

CMeEE

The evaluation task is the recognition of the named entity on the medical text. Given schema data and medical sentences, models are expected to extract entity about clinical information and classify these entities exactly.

CMeIE

The evaluation task is the extraction of entity relation on the medical text. Given schema and medical sentences, models are expected to automatically extract triples=[(S1, P1, O1), (S2, P2, O2)…] satisfying the constraint of schema. The schema defines the category of the predicate and corresponding subject and object, e.g.

（“subject_type”:“疾病”，“predicate”: “药物治疗”，“object_type”:“药物”） （“subject_type”:“疾病”，“predicate”: “实验室检查”，“object_type”:“检查”）

CHIP-CDN

The evaluation task is the normalization of the diagnosis entity from the Chinese medical record. Given a diagnosis entity, models are expected to return corresponding standard terms.

CHIP-CTC

In this evaluation task, given 44 semantic categories of screening standard (more detail in category.xlsx) and some description about Chinese clinical screening standard, models are expected to return every description's specific category.

CHIP-STS

In this evaluation task, given pairs of sentences involving five different diseases, models are expected to judge the semantic similarity of the pair of sentences.

KUAKE-QIC

In this evaluation task, given a medical query, models are expected to classify the intention of patients. These medical queries have 11 categories: diagnosis, cause, method, advice, metric explain, disease expression, result, attention, effect, price, other.

KUAKE-QTR

In this evaluation task, given a pair of query and title, models are expected to predict whether the topic of the pair query and title is consistent and the extent of their consistency.

KUAKE-QQR

In this evaluation task, given a pair of queries, models are expected to predict the extent of similarity between them.

Quick start

The modules of Data Processor, Model trainer could be found in cblue/. You can easily construct your code, train and evaluate your own models and methods. The corresponding Data Processor, Dataset, Trainer of eight tasks are listed below:

Example for CMeEE

from cblue.data import EEDataProcessor, EEDataset
from cblue.trainer import EETrainer
from cblue.metrics import ee_metric, ee_commit_prediction

# get samples
data_processor = EEDataProcessor(root=...)
train_samples = data_processor.get_train_sample()
eval_samples = data_processor.get_dev_sample()
test_samples = data_processor,get_test_sample()

# 'torch.Dataset'
train_dataset = EEDataset(train_sample, tokenizer=..., mode='train', max_length=...)

# training model
trainer = EETrainer(...)
trainer.train(...)

# predicton and generation of result
test_dataset = EEDataset(test_sample, tokenizer=..., mode='test', max_length=...)
trainer.predict(test_dataset)

Training setup

We list the hyper-parameters of every tasks during the baseline experiments.

Common hyper-parameters

CMeEE

Hyper-parameters for the training of pre-trained models with a token classification head on top for named entity recognition of the CMeEE task.

CMeIE-ER

Hyper-parameters for the training of pre-trained models with a token-level classifier for subject and object recognition of the CMeIE task.

CMeIE-RE

Hyper-parameters for the training of pre-trained models with a classifier for the entity pairs relation prediction of the CMeIE task.

CHIP-CTC

Hyper-parameters for the training of pre-trained models with a sequence classification head on top for screening criteria classification of the CHIP-CTC task.

CHIP-CDN-cls

Hyper-parameters for the CHIP-CDN task. We model the CHIP-CDN task with two stages: recall stage and ranking stage. num_negative_sample sets the number of negative samples sampled for the training ranking model during the ranking stage. recall_k sets the number of candidates recalled in the recall stage.

Hyper-parameters for the training of pre-trained models with a sequence classifier for the ranking model of the CHIP-CDN task. We encode the pairs of the original term and standard phrase from candidates recalled during the recall stage and then pass the pooled output to the classifier, which predicts the relevance between the original term and standard phrase.

CHIP-CDN-num

Hyper-parameters for the training of pre-trained models with a sequence classifier for the prediction of the number of standard phrases corresponding to the original term in the CHIP-CDN task. We take the prediction results of the model as the number we choose from the most relevant standard phrases, combining with the prediction of the ranking model.

CHIP-STS

Hyper-parameters for the training of pre-trained models with a sequence classifier for sentence similarity predication of the CHIP-STS task.

KUAKE-QIC

Hyper-parameters for the training of pre-trained models with a sequence classifier for query intention prediction of the KUAKE-QIC task.

KUAKE-QTR

Hyper-parameters for the training of pre-trained models with a sequence classifier for query-title pairs relevance prediction of the KUAKE-QTR task.

KUAKE-QQR

Hyper-parameters for the training of pre-trained models with a sequence classifier for query-query pairs relevance prediction of the KUAKE-QQR task.

How to Cite

@inproceedings{zhang-etal-2022-cblue,
    title = "{CBLUE}: A {C}hinese Biomedical Language Understanding Evaluation Benchmark",
    author = "Zhang, Ningyu  and
      Chen, Mosha  and
      Bi, Zhen  and
      Liang, Xiaozhuan  and
      Li, Lei  and
      Shang, Xin  and
      Yin, Kangping  and
      Tan, Chuanqi  and
      Xu, Jian  and
      Huang, Fei  and
      Si, Luo  and
      Ni, Yuan  and
      Xie, Guotong  and
      Sui, Zhifang  and
      Chang, Baobao  and
      Zong, Hui  and
      Yuan, Zheng  and
      Li, Linfeng  and
      Yan, Jun  and
      Zan, Hongying  and
      Zhang, Kunli  and
      Tang, Buzhou  and
      Chen, Qingcai",
    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.544",
    pages = "7888--7915",
    abstract = "Artificial Intelligence (AI), along with the recent progress in biomedical language understanding, is gradually offering great promise for medical practice. With the development of biomedical language understanding benchmarks, AI applications are widely used in the medical field. However, most benchmarks are limited to English, which makes it challenging to replicate many of the successes in English for other languages. To facilitate research in this direction, we collect real-world biomedical data and present the first Chinese Biomedical Language Understanding Evaluation (CBLUE) benchmark: a collection of natural language understanding tasks including named entity recognition, information extraction, clinical diagnosis normalization, single-sentence/sentence-pair classification, and an associated online platform for model evaluation, comparison, and analysis. To establish evaluation on these tasks, we report empirical results with the current 11 pre-trained Chinese models, and experimental results show that state-of-the-art neural models perform by far worse than the human ceiling.",
}

References

[1] CLUE: A Chinese Language Understanding Evaluation Benchmark [pdf] [git] [web]

[2] GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding [pdf] [web]

[3] SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems [pdf] [web]