/ReaLiSe

A Multi-modal Model Chinese Spell Checker Released on ACL2021.

Primary LanguagePythonMIT LicenseMIT

ReaLiSe

ReaLiSe is a multi-modal Chinese spell checking model.

This the office code for the paper Read, Listen, and See: Leveraging Multimodal Information Helps Chinese Spell Checking.

The paper has been accepted in ACL Findings 2021.

Environment

  • Python: 3.6
  • Cuda: 10.0
  • Packages: pip install -r requirements.txt

Data

Raw Data

SIGHAN Bake-off 2013: http://ir.itc.ntnu.edu.tw/lre/sighan7csc.html
SIGHAN Bake-off 2014: http://ir.itc.ntnu.edu.tw/lre/clp14csc.html
SIGHAN Bake-off 2015: http://ir.itc.ntnu.edu.tw/lre/sighan8csc.html
Wang271K: https://github.com/wdimmy/Automatic-Corpus-Generation

Data Processing

The code and cleaned data are in the data_process directory.

You can also directly download the processed data from this and put them in the data directory. The data directory would look like this:

data
|- trainall.times2.pkl
|- test.sighan15.pkl
|- test.sighan15.lbl.tsv
|- test.sighan14.pkl
|- test.sighan14.lbl.tsv
|- test.sighan13.pkl
|- test.sighan13.lbl.tsv

Pretrain

  • BERT: chinese-roberta-wwm-ext

    Huggingface hfl/chinese-roberta-wwm-ext: https://huggingface.co/hfl/chinese-roberta-wwm-ext
    Local: /data/dobby_ceph_ir/neutrali/pretrained_models/roberta-base-ch-for-csc/

  • Phonetic Encoder: pretrain_pho.sh

  • Graphic Encoder: pretrain_res.sh

  • Merge: merge.py

You can also directly download the pretrained and merged BERT, Phonetic Encoder, and Graphic Encoder from this, and put them in the pretrained directory:

pretrained
|- pytorch_model.bin
|- vocab.txt
|- config.json

Train

After preparing the data and pretrained model, you can train ReaLiSe by executing the train.sh script. Note that you should set up the PRETRAINED_DIR, DATE_DIR, and OUTPUT_DIR in it.

sh train.sh

Test

Test ReaLiSe using the test.sh script. You should set up the DATE_DIR, CKPT_DIR, and OUTPUT_DIR in it. CKPT_DIR is the OUTPUT_DIR of the training process.

sh test.sh

Well-trained Model

You can also download well-trained model from this direct using. The performance scores of RealiSe and some baseline models on the SIGHAN13, SIGHAN14, SIGHAN15 test set are here:

Methods

Metrics

  • "D" means "Detection Level", "C" means "Correction Level".
  • "A", "P", "R", "F" means "Accuracy", "Precision", "Recall", and "F1" respectively.

SIGHAN15

Method D-A D-P D-R D-F C-A C-P C-R C-F
FASpell 74.2 67.6 60.0 63.5 73.7 66.6 59.1 62.6
Soft-Masked BERT 80.9 73.7 73.2 73.5 77.4 66.7 66.2 66.4
SpellGCN - 74.8 80.7 77.7 - 72.1 77.7 75.9
BERT 82.4 74.2 78.0 76.1 81.0 71.6 75.3 73.4
ReaLiSe 84.7 77.3 81.3 79.3 84.0 75.9 79.9 77.8

SIGHAN14

Method D-A D-P D-R D-F C-A C-P C-R C-F
Pointer Network - 63.2 82.5 71.6 - 79.3 68.9 73.7
SpellGCN - 65.1 69.5 67.2 - 63.1 67.2 65.3
BERT 75.7 64.5 68.6 66.5 74.6 62.4 66.3 64.3
ReaLiSe 78.4 67.8 71.5 69.6 77.7 66.3 70.0 68.1

SIGHAN13

Method D-A D-P D-R D-F C-A C-P C-R C-F
FASpell 63.1 76.2 63.2 69.1 60.5 73.1 60.5 66.2
SpellGCN 78.8 85.7 78.8 82.1 77.8 84.6 77.8 81.0
BERT 77.0 85.0 77.0 80.8 77.4 83.0 75.2 78.9
ReaLiSe 82.7 88.6 82.5 85.4 81.4 87.2 81.2 84.1

Citation

@misc{xu2021read,
      title={Read, Listen, and See: Leveraging Multimodal Information Helps Chinese Spell Checking}, 
      author={Heng-Da Xu and Zhongli Li and Qingyu Zhou and Chao Li and Zizhen Wang and Yunbo Cao and Heyan Huang and Xian-Ling Mao},
      year={2021},
      eprint={2105.12306},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}