Skip to content

Latest commit

 

History

History
73 lines (63 loc) · 2.06 KB

README.md

File metadata and controls

73 lines (63 loc) · 2.06 KB
Raw

CMRC-2018 task

CMRC-2018 is a machine reading comprehension task.

Dataset

The corpus can be find HERE
Download the corpus and save data at [CMRC_DATA_PATH]

Model

Download ChineseBERT model and save at [CHINESEBERT_PATH].

Preprocess

For MRC tasks, the length of document usually larger than 512, so we have to do some pre-process, such as sliding window.
Run the following script to process training data and save the result at [CMRC_TRAIN_PATH]

python data_generate.py \
 --bert_path [CHINESEBERT_PATH] \
 --data_dir [CMRC_DATA_PATH] \
 --output_dir  [CMRC_TRAIN_PATH]

Train and Evaluate

Run the following scripts to generate test file .

python cmrc_trainer.py \
      --bert_path [CHINESEBERT_PATH] \
      --data_dir [CMRC_TRAIN_PATH] \
      --save_path [OUTPUT_PATH] \
      --gpus=0,1,2,3 \
      --batch_size=8 \
      --lr=3e-5 \
      --max_epoch=2 \
      --val_check_interval 0.1 \
      --accumulate_grad_batches=4 \
      --warmup_proporation 0.1

the final checkpoint will be saved at [OUTPUT_PATH]/checkpoint, run following script to generate test file

python cmrc_evluate.py \
    --bert_path [CHINESEBERT_PATH]  \
    --save_path [OUTPUT_PATH]  \
    --test_file [CMRC_DATA_PATH]/test.json  \
    --pretrain_checkpoint [OUTPUT_PATH]/checkpoint/****.ckpt \
    --gpus=0,

The test file will be generated at [OUTPUT_PATH]\test_predictions.json, upload the file to CLUE to get the test score.

Result

The evaluation metric is EM.
Result of our model and previous SOTAs are:

base model:

Model Dev Test
ERNIE 66.89 74.70
BERT 66.77 71.60
BERT-wwm 66.96 73.95
RoBERTa 67.89 75.20
MacBERT - -
ChineseBERT 67.95 95.7

large model:

Model Dev Test
RoBERTa 70.59 77.95
MacBERT - -
ChineseBERT 70.70 78.05