A Chinese Electric Health Record Bert Pretrained Model.
The cleaner is responsible for cleaning txt files, which is used for training a Chinese bert model. The cleaner split lines in original lines into small lines. Each small line is a complete sentence with a punctuation. This is required for training next sentence prediction task.
cd ./cleaner/ python parser.py [-h] [--input INPUT] [--output OUTPUT] [-s] [--log LOG]
--input: input directory --output: output directory -s: output is one single file --log: log frequency train
Pre-train a bert model with cleaned text. We should generate .tfrecord first, and pre-train with google's code. To notice, cleaner file may be too big to load in RAM. Our script splits these files and generate multiple .tfrecord.
Split file and convert to .tfrecord
cd ./train/ python make_pretrain_bert.py [-h] [-f FILE_PATH] [-s SPLIT_LINE] [-p SPLIT_PATH] [-o OUTPUT_PATH] [-l MAX_LENGTH] [-b BERT_BASE_DIR]
-f: cleaned file path -s: split line count, default=500000 -p: splited file save path -o: .tfrecord save path -l: max length -b: bert base dir One should change parameters for your specific requirement in pretrain128.sh and pretrain512.sh .
sh pretrain128.sh sh pretrain512.sh
Test Chinese medical NLP tasks by BERT in one line! Two NER tasks, one QA task, one RE task and one sentence similarity task.
cd ./test/ sh run_test.sh
Tasks include CCKS2019NER , cMedQA2 , Tianchi_NER , Tianchi_RE , ncov2019_sim .
Results compared with original BERT and ChineseEHRBert. Results are preparing.
Zheng Yuan Peng Zhao Chen Yu Sheng Yu