macbert4mdcspell_v2

macbert4mdcspell

概述(macbert4mdcspell)

• macro-correct, 中文拼写纠错CSC测评(文本纠错), 权重使用
• 项目地址在 https://github.com/yongzhuo/macro-correct
• 本模型权重为macbert4mdcspell_v2, 使用mdcspell架构, 其特点是det_label和cor_label交互;
• 训练时加入了macbert的mlm-loss, 推理时舍弃了macbert后面的部分;
• 如何使用: 1.使用transformers调用; 2.使用macro-correct项目调用; 详情见三、调用(Usage);
• 为了修复过纠问题, macbert4mdcspell_v2的MFT只70%的时间no-error-mask(0.15), 15%的时间target-to-target, 15%的时间不mask;

目录

• 一、测评(Test)
• 二、结论(Conclusion)
• 三、调用(Usage)
• 四、论文(Paper)
• 五、参考(Refer)
• 六、引用(Cite)

一、测评(Test)

1.1 测评数据来源

地址为Macropodus/csc_eval_public, 所有训练数据均来自公网或开源数据, 训练数据为1千万左右, 混淆词典较大;

1.gen_de3.json(5545): '的地得'纠错, 由人民日报/学习强国/chinese-poetry等高质量数据人工生成;
2.lemon_v2.tet.json(1053): relm论文提出的数据, 多领域拼写纠错数据集(7个领域), ; 包括game(GAM), encyclopedia (ENC), contract (COT), medical care(MEC), car (CAR), novel (NOV), and news (NEW)等领域;
3.acc_rmrb.tet.json(4636): 来自NER-199801(人民日报高质量语料);
4.acc_xxqg.tet.json(5000): 来自学习强国网站的高质量语料;
5.gen_passage.tet.json(10000): 源数据为qwen生成的好词好句, 由几乎所有的开源数据汇总的混淆词典生成;
6.textproof.tet.json(1447): NLP竞赛数据, TextProofreadingCompetition;
7.gen_xxqg.tet.json(5000): 源数据为学习强国网站的高质量语料, 由几乎所有的开源数据汇总的混淆词典生成;
8.faspell.dev.json(1000): 视频字幕通过OCR后获取的数据集; 来自爱奇艺的论文faspell;
9.lomo_tet.json(5000): 主要为音似中文拼写纠错数据集; 来自腾讯; 人工标注的数据集CSCD-NS;
10.mcsc_tet.5000.json(5000): 医学拼写纠错; 来自腾讯医典APP的真实历史日志; 注意论文说该数据集只关注医学实体的纠错, 常用字等的纠错并不关注;
11.ecspell.dev.json(1500): 来自ECSpell论文, 包括(law/med/gov)等三个领域;
12.sighan2013.dev.json(1000): 来自sighan13会议;
13.sighan2014.dev.json(1062): 来自sighan14会议;
14.sighan2015.dev.json(1100): 来自sighan15会议;

1.2 测评数据预处理

测评数据都经过 全角转半角,繁简转化,标点符号标准化等操作;

1.3 其他说明

1.指标带common的极为宽松指标, 同开源项目pycorrector的评估指标;
2.指标带strict的极为严格指标, 同开源项目[wangwang110/CSC](https://github.com/wangwang110/CSC);
3.macbert4mdcspell_v1模型为训练使用mdcspell架构+bert的mlm-loss, 但是推理的时候只用bert-mlm;
4.acc_rmrb/acc_xxqg数据集没有错误, 用于评估模型的误纠率(过度纠错);
5.qwen25_1-5b_pycorrector的模型为shibing624/chinese-text-correction-1.5b, 其训练数据包括了lemon_v2/mcsc_tet/ecspell的验证集和测试集, 其他的bert类模型的训练不包括验证集和测试集;

二、重要指标

2.1 F1(common_cor_f1)

2.2 acc(common_cor_acc)

2.3 acc(acc_true, thr=0.75)

二、结论(Conclusion)

1.macbert4csc_v1/macbert4csc_v2/macbert4mdcspell_v1等模型使用多种领域数据训练, 比较均衡, 也适合作为第一步的预训练模型, 可用于专有领域数据的继续微调;
2.比较macbert4csc_pycorrector/bertbase4csc_v1/macbert4csc_v2/macbert4mdcspell_v1, 观察表2.3, 可以发现训练数据越多, 准确率提升的同时, 误纠率也会稍微高一些;
3.MFT(Mask-Correct)依旧有效, 不过对于数据量足够的情形提升不明显, 可能也是误纠率升高的一个重要原因;
4.训练数据中也存在文言文数据, 训练好的模型也支持文言文纠错;
5.训练好的模型对"地得的"等高频错误具有较高的识别率和纠错率;
6.macbert4mdcspell_v2的MFT只70%的时间no-error-mask(0.15), 15%的时间target-to-target, 15%的时间不mask;

三、调用(Usage)

3.1 使用macro-correct

import os
os.environ["MACRO_CORRECT_FLAG_CSC_TOKEN"] = "1"
from macro_correct import correct
### 默认纠错(list输入)
text_list = ["真麻烦你了。希望你们好好的跳无",
             "少先队员因该为老人让坐",
             "机七学习是人工智能领遇最能体现智能的一个分知",
             "一只小鱼船浮在平净的河面上"
             ]
text_csc = correct(text_list)
print("默认纠错(list输入):")
for res_i in text_csc:
    print(res_i)
print("#" * 128)

"""
默认纠错(list输入):
{'index': 0, 'source': '真麻烦你了。希望你们好好的跳无', 'target': '真麻烦你了。希望你们好好地跳舞', 'errors': [['的', '地', 12, 0.6584], ['无', '舞', 14, 1.0]]}
{'index': 1, 'source': '少先队员因该为老人让坐', 'target': '少先队员应该为老人让坐', 'errors': [['因', '应', 4, 0.995]]}
{'index': 2, 'source': '机七学习是人工智能领遇最能体现智能的一个分知', 'target': '机器学习是人工智能领域最能体现智能的一个分支', 'errors': [['七', '器', 1, 0.9998], ['遇', '域', 10, 0.9999], ['知', '支', 21, 1.0]]}
{'index': 3, 'source': '一只小鱼船浮在平净的河面上', 'target': '一只小鱼船浮在平静的河面上', 'errors': [['净', '静', 8, 0.9961]]}
"""

3.2 使用 transformers

# !/usr/bin/python
# -*- coding: utf-8 -*-
# @time    : 2021/2/29 21:41
# @author  : Mo
# @function: transformers直接加载bert类模型测试


import traceback
import time
import sys
import os
os.environ["USE_TORCH"] = "1"
from transformers import BertConfig, BertTokenizer, BertForMaskedLM
import torch

# pretrained_model_name_or_path = "shibing624/macbert4csc-base-chinese"
pretrained_model_name_or_path = "Macropodus/macbert4mdcspell_v2"
# pretrained_model_name_or_path = "Macropodus/macbert4mdcspell_v1"
# pretrained_model_name_or_path = "Macropodus/macbert4csc_v1"
# pretrained_model_name_or_path = "Macropodus/macbert4csc_v2"
# pretrained_model_name_or_path = "Macropodus/bert4csc_v1"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
max_len = 128

print("load model, please wait a few minute!")
tokenizer = BertTokenizer.from_pretrained(pretrained_model_name_or_path)
bert_config = BertConfig.from_pretrained(pretrained_model_name_or_path)
model = BertForMaskedLM.from_pretrained(pretrained_model_name_or_path)
model.to(device)
print("load model success!")

texts = [
    "机七学习是人工智能领遇最能体现智能的一个分知",
    "我是练习时长两念半的鸽仁练习生蔡徐坤",
    "真麻烦你了。希望你们好好的跳无",
    "他法语说的很好，的语也不错",
    "遇到一位很棒的奴生跟我疗天",
    "我们为这个目标努力不解",
]
len_mid = min(max_len, max([len(t)+2 for t in texts]))

with torch.no_grad():
    outputs = model(**tokenizer(texts, padding=True, max_length=len_mid,
                                return_tensors="pt").to(device))

def get_errors(source, target):
    """   极简方法获取 errors   """
    len_min = min(len(source), len(target))
    errors = []
    for idx in range(len_min):
        if source[idx] != target[idx]:
            errors.append([source[idx], target[idx], idx])
    return errors

result = []
for probs, source in zip(outputs.logits, texts):
    ids = torch.argmax(probs, dim=-1)
    tokens_space = tokenizer.decode(ids[1:-1], skip_special_tokens=False)
    text_new = tokens_space.replace(" ", "")
    target = text_new[:len(source)]
    errors = get_errors(source, target)
    print(source, " => ", target, errors)
    result.append([target, errors])
print(result)
"""
机七学习是人工智能领遇最能体现智能的一个分知  =>  机器学习是人工智能领域最能体现智能的一个分支 [['七', '器', 1], ['遇', '域', 10], ['知', '支', 21]]
我是练习时长两念半的鸽仁练习生蔡徐坤  =>  我是练习时长两年半的个人练习生蔡徐坤 [['念', '年', 7], ['鸽', '个', 10], ['仁', '人', 11]]
真麻烦你了。希望你们好好的跳无  =>  真麻烦你了。希望你们好好地跳舞 [['的', '地', 12], ['无', '舞', 14]]
他法语说的很好，的语也不错  =>  他法语说得很好，德语也不错 [['的', '得', 4], ['的', '德', 8]]
遇到一位很棒的奴生跟我疗天  =>  遇到一位很棒的女生跟我聊天 [['奴', '女', 7], ['疗', '聊', 11]]
我们为这个目标努力不解  =>  我们为这个目标努力不懈 [['解', '懈', 10]]
"""

四、论文(Paper)

• 2024-Refining: Refining Corpora from a Model Calibration Perspective for Chinese

• 2024-ReLM: Chinese Spelling Correction as Rephrasing Language Model

• 2024-DICS: DISC: Plug-and-Play Decoding Intervention with Similarity of Characters for Chinese Spelling Check

• 2023-Bi-DCSpell: A Bi-directional Detector-Corrector Interactive Framework for Chinese Spelling Check

• 2023-BERT-MFT: Rethinking Masked Language Modeling for Chinese Spelling Correction

• 2023-PTCSpell: PTCSpell: Pre-trained Corrector Based on Character Shape and Pinyin for Chinese Spelling Correction

• 2023-DR-CSC: A Frustratingly Easy Plug-and-Play Detection-and-Reasoning Module for Chinese

• 2023-DROM: Disentangled Phonetic Representation for Chinese Spelling Correction

• 2023-EGCM: An Error-Guided Correction Model for Chinese Spelling Error Correction

• 2023-IGPI: Investigating Glyph-Phonetic Information for Chinese Spell Checking: What Works and What’s Next?

• 2023-CL: Contextual Similarity is More Valuable than Character Similarity-An Empirical Study for Chinese Spell Checking

• 2022-CRASpell: CRASpell: A Contextual Typo Robust Approach to Improve Chinese Spelling Correction

• 2022-MDCSpell: MDCSpell: A Multi-task Detector-Corrector Framework for Chinese Spelling Correction

• 2022-SCOPE: Improving Chinese Spelling Check by Character Pronunciation Prediction: The Effects of Adaptivity and Granularity

• 2022-ECOPO: The Past Mistake is the Future Wisdom: Error-driven Contrastive Probability Optimization for Chinese Spell Checking

• 2021-MLMPhonetics: Correcting Chinese Spelling Errors with Phonetic Pre-training

• 2021-ChineseBERT: ChineseBERT: Chinese Pretraining Enhanced by Glyph and Pinyin Information

• 2021-BERTCrsGad: Global Attention Decoder for Chinese Spelling Error Correction

• 2021-ThinkTwice: Think Twice: A Post-Processing Approach for the Chinese Spelling Error Correction

• 2021-PHMOSpell: PHMOSpell: Phonological and Morphological Knowledge Guided Chinese Spelling Chec

• 2021-SpellBERT: SpellBERT: A Lightweight Pretrained Model for Chinese Spelling Check

• 2021-TwoWays: Exploration and Exploitation: Two Ways to Improve Chinese Spelling Correction Models

• 2021-ReaLiSe: Read, Listen, and See: Leveraging Multimodal Information Helps Chinese Spell Checking

• 2021-DCSpell: DCSpell: A Detector-Corrector Framework for Chinese Spelling Error Correction

• 2021-PLOME: PLOME: Pre-training with Misspelled Knowledge for Chinese Spelling Correction

• 2021-DCN: Dynamic Connected Networks for Chinese Spelling Check

• 2020-SoftMaskBERT: Spelling Error Correction with Soft-Masked BERT

• 2020-SpellGCN: SpellGCN：Incorporating Phonological and Visual Similarities into Language Models for Chinese Spelling Check

• 2020-ChunkCSC: Chunk-based Chinese Spelling Check with Global Optimization

• 2020-MacBERT: Revisiting Pre-Trained Models for Chinese Natural Language Processing

• 2019-FASPell: FASPell: A Fast, Adaptable, Simple, Powerful Chinese Spell Checker Based On DAE-Decoder Paradigm

• 2018-Hybrid: A Hybrid Approach to Automatic Corpus Generation for Chinese Spelling Checking

• 2015-Sighan15: Introduction to SIGHAN 2015 Bake-off for Chinese Spelling Check

• 2014-Sighan14: Overview of SIGHAN 2014 Bake-off for Chinese Spelling Check

• 2013-Sighan13: Chinese Spelling Check Evaluation at SIGHAN Bake-off 2013

五、参考(Refer)

• nghuyong/Chinese-text-correction-papers
• destwang/CTCResources
• wangwang110/CSC
• chinese-poetry/chinese-poetry
• chinese-poetry/huajianji
• garychowcmu/daizhigev20
• yangjianxin1/Firefly
• Macropodus/xuexiqiangguo_428w
• Macropodus/csc_clean_wang271k
• Macropodus/csc_eval_public
• shibing624/pycorrector
• iioSnail/MDCSpell_pytorch
• gingasan/lemon
• Claude-Liu/ReLM

六、引用(Cite)

For citing this work, you can refer to the present GitHub project. For example, with BibTeX:

@software{macro-correct,
    url = {https://github.com/yongzhuo/macro-correct},
    author = {Yongzhuo Mo},
    title = {macro-correct},
    year = {2025}