This is the official repo for the ACL 2023 paper Rethinking Masked Language Modeling for Chinese Spelling Correction; AAAI 2024 paper Chinese Spelling Correction as Rephraing Language Model.

Fine-tuning results on some of benchmarks:

==New==

ReLM

ReLM pre-trained model is released. It is a rephrasing language model trained based on bert-base-chinese and 34 million monolingual data.

relm-m0.2.bin

relm-m0.3.bin

Different from BERT-MFT, ReLM is a pure language model, which optimizes the rephrasing language modeling objective instead of sequence tagging.

from autocsc import AutoCSCReLM

model = AutoCSCReLM.from_pretrained("bert-base-chinese",
                                    state_dict=torch.load("relm-m0.3.bin"),
                                    cache_dir="cache")

Monolingual data

We share our used training data for LEMON. It contains 34 million monolingual sentences and we synthesize sentence pairs based on our confusion set in confus.

monolingual-wiki-news-l64

We split the data into 343 sub-files with 100,000 sentences for each. The total size of the .zip file is 1.5G.

Our code supports multiple GPUs now:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 accelerate launch --multi_gpu run.py \
  --do_train \
  --do_eval \
  --fp16 \
  --mft

LEMON (large-scale multi-domain dataset with natural spelling errors) is a novel benchmark released with our paper. All test sets are in lemon_v2.

Note: This dataset can only be used for academic research, it cannot be used for commercial purposes.

The other test sets we use in the paper are in sighan_ecspell.

The confusion sets are in confus.

Trained weights

In our paper, we train BERT for 30,000 steps, with the learning rate 5e-5 and batch size 8192. The backbone model is bert-base-chinese. We share our trained model weights to facilitate future research. We welcome researchers to develop better ones based on our models.

BERT-finetune-MFT

BERT-finetune-MFT-CreAT-maskany

BERT-SoftMasked-MFT

We implement some architectures in recent CSC papers in autocsc.py.

• Spelling Error Correction with Soft-Masked BERT

• MDCSpell: A Multi-task Detector-Corrector Framework for Chinese Spelling Correction

For instance (Soft-Masked BERT):

from autocsc import AutoCSCSoftMasked

# Load the model, similar to huggingface transformers.
model = AutoCSCSoftMasked.from_pretrained("bert-base-chinese",
                                          cache_dir="cache")

# Go forward step.
outputs = model(src_ids=src_ids,
                attention_mask=attention_mask,
                trg_ids=trg_ids)
loss = outputs["loss"]
prd_ids = outputs["predict_ids"].tolist()

If you have new models or suggestions for promoting our implementations, feel free to email me.

Running (set --mft for Masked-FT):

CUDA_VISIBLE_DEVICES=0 python run.py \
  --do_train \
  --do_eval \
  --train_on xxx.txt \
  --eval_on xx.txt \
  --output_dir mft \
  --max_train_steps 10000 \
  --fp16 \
  --model_type mdcspell \
  --mft

Directly testing on LEMON (including SIGHAN):

CUDA_VISIBLE_DEVICES=0 python run.py \
  --test_on_lemon ../data/lemon \
  --output_dir relm \
  --model_type relm \
  --load_state_dict relm-m0.3.bin