Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Mixture of Small and Large Models for Chinese Spelling Check

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Mixing a fine-tuned classifier's per-character scores into a frozen LLM's beam search produces state-of-the-art Chinese spelling correction without LLM fine-tuning.

desk verdict Solid empirical CSC paper with a genuinely new decoding-time fusion, but the headline SOTA numbers rest on test-set-tuned weights; worth reviewing, needs a held-out protocol. read the letter →

arxiv 2506.06887 v1 pith:ZB2UNG3M submitted 2025-06-07 cs.CL

Ziheng Qiao , Houquan Zhou , Zhenghua Li This is my paper
classification cs.CL
keywords ChinesespellingcheckmixtureofsmallandlargemodelsbeamsearchdecodinglanguageBERT-basedclassificationdistortionmodelentropyweightingdomainadaptation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Chinese spelling check (CSC) can be done by fine-tuned character classifiers, which are accurate but overfit to edit patterns they saw in training, or by large language models used as pure language models, which generalize across domains but miss many errors. This paper's central claim is that the two can be combined at inference time by adding the small model's per-character probabilities into the LLM's beam-search score, with the LLM left completely frozen. The resulting mixture outperforms either component alone and reaches new state-of-the-art results on five public benchmarks. Because only the small model needs fine-tuning, domain adaptation becomes cheap, and the paper demonstrates that the same fixed-weight recipe transfers across three open-source LLM families and sizes from 0.5B to 20B parameters.

What carries the argument

The central object is the incremental beam-search score of Eq. (1): each new token adds $\log p_{\mathrm{LLM}}(t_k \mid t_{<k})$ plus $(1 + H_{\mathrm{LLM}}(\cdot))$ times $\alpha \log p_{\mathrm{DM}}(x,i \mid t_k) + \beta \log p_{\mathrm{SM}}(t_k \mid x,i)$, where $H_{\mathrm{LLM}}$ is the LLM's entropy on the token distribution. The distortion model $p_{\mathrm{DM}}$ assigns fixed probabilities by character-pair type (identical, same pinyin, similar pinyin, similar shape, unrelated), inherited from the prior distortion-model work; the small-model term $p_{\mathrm{SM}}$ multiplies per-character softmax probabilities from a fine-tuned BERT-style classifier. This object carries the argument because it is the only point of contact between the two models: it converts the small model's local, position-wise judgement into a global sequence score that the LLM's beam search can optimize, while the entropy factor lets the LLM decide when to rely on the other two components.

What would settle it

Take a new native-speaker domain not used in the paper, run the mixture with the fixed reported weights $\alpha=0.5$, $\beta=0.9$, and compare against the better of the small model alone and the distortion-model LLM alone; the central claim would be undercut if the mixture does not beat both, or if the optimal weights shift enough across domains that no fixed $\alpha,\beta$ works.

Watch

Extended reading notes

Core claim

The central claim is that a weighted sum of three log-probabilities in Eq. (1) — the LLM's autoregressive token probability, a distortion-model probability enforcing faithfulness between input and output characters, and the small fine-tuned classifier's per-character softmax probability — outperforms both the small model alone and the distortion-model LLM approach alone. The small-model term is the product of per-character probabilities over the characters spanned by each LLM token, so token-level alignment is handled by splitting LLM tokens into characters. The weight of the non-LLM terms is modulated by the LLM's entropy, so the classifier and distortion model are heard more when the LLM is uncertain. With this recipe the paper reports state-of-the-art correction F1 on rSIGHAN15, CSCD-NS, MCSCSet, ECSpell, and LEMON, and it shows the gain is robust across LLM family, model size, beam size, and small-model choices.

Load-bearing premise

The assumption that carries the method is that the LLM's token log-probability, the distortion model's log-probability, and the fine-tuned classifier's per-character log-probabilities are on comparable scales, so that fixed weights $\alpha$ and $\beta$ can blend them; the paper selects these weights by grid search on a few datasets and does not calibrate the scales across LLM families or domains.

Editorial extensions

If this is right

  • Any fine-tuned character classifier can be plugged into a frozen LLM's beam search with token-level alignment, making the recipe model-agnostic on both sides.
  • Fine-tuning the LLM is unnecessary for the reported gains; the cost is fine-tuning a small model plus one extra forward pass of it per beam step.
  • The entropy term means the mixture automatically leans on the classifier in uncertain LLM contexts and lets the LLM dominate fluent ones.
  • Even without the distortion model and faithfulness reward, the mixture still beats the small model alone, locating the core gain in LLM language modelling plus the classifier's corrections.
  • The fixed weights $\alpha=0.5$, $\beta=0.9$ are a single operating point; the paper's own curves show tuning them can raise performance further.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same score fusion should transfer to other equal-length editing tasks such as grammatical error correction, provided the token alignment between classifier and LLM is handled the same way; the paper lists this as a possible extension but does not test it.
  • Editorial inference: because the LLM stays frozen, per-domain adaptation reduces to fine-tuning the small classifier, which suggests a deployment pattern of one frozen LLM plus many cheap domain-specific classifiers; the paper's domain-adaptation discussion points there but stops short of building it.
  • Editorial inference: the ECSpell leakage finding implies some published small-model numbers on that benchmark are inflated, so mixtures that include a frozen LLM may be the fairer cross-domain estimate; checking benchmark splits for repeated source sentences would be a cheap routine safeguard.
  • Editorial inference: the shift of the optimal $\beta$ across LLM families in Figure 3 suggests an automatic, confidence-based schedule for $\beta$ could replace manual grid search; the paper does not test this.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a mixture approach for Chinese Spelling Check (CSC) that combines a fine-tuned BERT-based small model with an open-source LLM at decoding time. The core scoring function (Eq. 1) adds the small model's per-character log-probability and the distortion-model log-probability, scaled by weights α and β, to the LLM's token log-probability, with an entropy-based faithfulness reward. The authors report state-of-the-art results on rSIGHAN15, CSCD-NS, MCSCSet, ECSpell, and LEMON, using three 7B LLM families (Baichuan2, Qwen2.5, InternLM2.5) and several small models. They also ablate the distortion model and faithfulness reward, analyze hyperparameter sensitivity, and provide a cleaned version of ECSpell after discovering target-sentence leakage.

Significance. If the reported results are reliable, the approach is simple, training-free for the LLM, and broadly applicable: it improves over both small models and LLM-only baselines across multiple domains. The paper includes extensive experiments, code release, and a careful re-evaluation of ECSpell. The main risk is that the key hyperparameters α and β appear to be selected on test benchmarks, which could inflate the SOTA claims. The reuse of the distortion table from Zhou et al. (2024) without recalibration also raises a scale-compatibility question for the additive score in Eq. (1). These issues are fixable and do not invalidate the core idea.

major comments (3)
  1. [Section 6.2 / Figure 3 / Table 2] The paper does not describe a held-out validation protocol for choosing α=0.5 and β=0.9 in Eq. (1). Figure 3 directly sweeps these weights on the test sets (rSIGHAN15, ECSpell-Odw, LEMON-Nov) and Section 6.2 states that 'tuning these weights can further enhance model performance.' This suggests the fixed weights used in Table 2 may have been selected with test-set knowledge, which would optimistically bias the SOTA comparisons. The authors should either adopt a proper validation split for hyperparameter selection, or demonstrate that the chosen weights are robust across a range of settings for all baselines, not only for the ReLM baseline as currently stated.
  2. [Section 3.2, Eq. (1)] The mixture score assumes that log p_LLM, log p_DM, and log p_SM are on comparable scales and can be added with fixed weights α and β. No calibration analysis is provided, and the distortion table (Table 1) is inherited from Zhou et al. (2024) without recalibration. Figure 3 shows that the optimal β shifts across LLM families and datasets (e.g., Qwen2.5 and InternLM2.5 require larger β on rSIGHAN15). This indicates that the components are not commensurable across settings, so the fixed-weight recipe may not transfer to new domains. The authors should analyze the score distributions or provide a principled calibration mechanism for α and β.
  3. [Section 5, Table 2 and Section 6.2] The robustness claim 'our approach consistently surpasses the ReLM baseline across all hyperparameter settings' only compares against ReLM. For the other baselines (BERT, ReaLiSe, SCOPE, etc.), results are reported solely at the chosen weights. If the weights are tuned on the test sets, the comparison to these baselines is not fair. The authors should show that the mixture improves over all baselines across a range of α and β, or at least for weights selected via a validation set, to support the SOTA claim.
minor comments (5)
  1. [Section 3.2] The term 'dynamic mixture' is somewhat misleading because only the entropy multiplier (1 + H_LLM(·)) is dynamic; α and β are fixed constants. Consider clarifying the terminology.
  2. [Section 6.2] The sentence 'In practice, tuning these weights can further enhance model performance' is in tension with the use of fixed weights in Table 2. It would be clearer to explicitly state that the fixed weights are a default and that per-domain tuning is possible, and then specify the protocol for that tuning.
  3. [Table 4] The effect of removing the faithfulness reward (FR) is mixed across datasets: positive on ECSpell-Odw (+0.4 S-F) but negative on rSIGHAN15 (-2.1 S-F) and neutral on LEMON-Nov. This is not discussed in the text; please comment on why the FR helps only on some datasets.
  4. [Appendix C.3] The ECSpell cleaning is a strength, but the percentages of overlapping sentences (52.7%, 19.3%, 28.2%) are stated without describing the matching criterion (e.g., exact source sentence match, reference sentence match, or both). Please clarify the procedure.
  5. [Tables 13-15] The domain-specific tables have many columns without clear headers or row groupings; consider adding explicit column labels and separators for readability.

Circularity Check

2 steps flagged · score 2.0 of 10

No constructional circularity: the mixture claim is measured on external benchmarks; only minor self-citation and a test-set hyperparameter selection concern are present.

  1. other [Section 6.2 (Impact of Hyperparameters), Figure 3, Table 2]
    "Figure 3: Model performance (sentence-level F1) of ReLM + LLMs on rSIGHAN15, ECSpell, and LEMON with different α and β. The x-axis is α and y-axis is β. ... It is important to note that in Table 2, the weights are fixed (α=0.5, β=0.9). In practice, tuning these weights can further enhance model performance."

    The SOTA results in Table 2 are produced with α=0.5 and β=0.9, while Section 6.2 and Figure 3 sweep these same weights on the same test benchmarks (rSIGHAN15, ECSpell-Odw, LEMON-Nov) without reporting a held-out validation split. The statement that tuning can further improve performance confirms the weights are not fixed a priori. Thus the headline gains are evaluated at hyperparameters selected on the test sets rather than at pre-specified or independently validated settings. This is a mild fitted-input concern, not a constructional reduction, because Eq. (1) is an explicit modeling ansatz and the benchmark numbers are external measurements.

  2. other [Section 3.2, Eq. (1), Table 1]
    "Moreover, we follow Zhou et al. (2024) and employ their proposed faithfulness reward to further encourage that y retains the same meaning as x. ... Table 1: The distribution of the different distortion types extracted from Zhou et al. (2024)."

    The distortion-model probabilities in Table 1 and the faithfulness reward H_LLM in Eq. (1) are imported from Zhou et al. (2024), a paper co-authored by Houquan Zhou and Zhenghua Li from the same group. These components are part of the score being evaluated, so part of the method is inherited from the authors' own prior work. However, this is not the paper's claimed contribution, and the ablation in Table 4 shows that removing DM and FR degrades but does not eliminate the improvement; the additive small-model term is independently evaluated against external benchmarks. Hence this is a minor self-citation issue rather than a load-bearing circularity.

full rationale

The paper's central claim is empirical: adding beta * log p_SM to the LLM-plus-distortion score in Eq. (1) and running beam search improves correction F1 on public benchmarks. This claim is tested against external, independently constructed test sets (rSIGHAN15, CSCD-NS, MCSCSet, ECSpell, LEMON), so it does not reduce by definition to the components of the score. Eq. (1) is an explicit design ansatz, not a theorem derived from premises that already contain the result. The two concerns are (a) the weights alpha=0.5 and beta=0.9 appear to be selected using sweeps on the same test sets, with no held-out split described, and (b) the distortion model and faithfulness reward are inherited from Zhou et al. (2024), which shares co-authors with this paper. Neither is constructional: the ablation shows the mixture remains above the ReLM baseline even without the inherited components, and the benchmark numbers are measurements rather than consequences of the ansatz. Score 2 reflects these minor issues; there is no circular derivation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The proposed score in Eq. (1) is a weighted sum of three probabilistic terms. The only new numbers contributed by this paper are the weights alpha and beta; the distortion table is reused from the authors' prior work, and the small model is an off-the-shelf fine-tuned classifier. The main modeling commitments are the per-position independence assumptions in p_SM and p_DM and the use of LLM entropy as a confidence signal. There are no invented physical or conceptual entities.

free parameters (3)
  • alpha (distortion model weight) = 0.5
    Fixed value used for all main results in Table 2. Figure 3 shows that the optimum varies with LLM family and dataset, so this is a hand-set or tuned hyperparameter rather than a derived constant.
  • beta (small model weight) = 0.9
    Fixed value used for all main results. Section 6.2 indicates that Qwen2.5 and InternLM2.5 prefer higher beta values, so the chosen value is tuned rather than derived.
  • Distortion-type probability table = identical 0.962; same pinyin 0.023; similar pinyin 0.008; similar shape 0.004; unrelated 0.003
    Inherited from Zhou et al. (2024) and used to compute p_DM in Eq. (1). These corpus-derived probabilities are not re-estimated on the test domains, and the mixture's faithfulness term depends on them.
assumptions (4)
  • domain assumption The small model's per-character decisions factor as p_SM(y|x) = product over i of p_SM(y_i|x,i).
    Used in Section 3 to define the small-model term. Each position is scored independently given the input, ignoring joint dependencies among corrected characters.
  • domain assumption The base LLM's next-token distribution is a usable fluency prior, and its entropy H_LLM is a usable confidence signal for re-weighting the auxiliary terms.
    Eq. (1) multiplies the small-model and distortion terms by (1+H_LLM). The paper assumes higher LLM uncertainty should boost the other components, without calibrating this heuristic.
  • domain assumption The distortion model and its fixed probability table from Zhou et al. (2024) transfer to all test domains.
    The faithfulness term p_DM uses Table 1 unchanged. Section 6.3 shows that removing this term lowers sentence-level F1 by up to 8 points on LEMON-Nov, so the assumption is load-bearing.
  • standard math Beam search over the mixture score approximates the global optimum of score(x,y).
    Standard decoding assumption. Section 6.7 checks beam-size sensitivity but does not prove convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mixture of Small and Large Models for Chinese Spelling Check." pith.science (2026). https://pith.science/paper/ZB2UNG3M

@misc{pith2026250606887,
  author       = {Pith},
  title        = {Pith review of: Mixture of Small and Large Models for Chinese Spelling Check},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZB2UNG3M}},
  note         = {Machine review of arXiv:2506.06887}
}
read the original abstract

In the era of large language models (LLMs), the Chinese Spelling Check (CSC) task has seen various LLM methods developed, yet their performance remains unsatisfactory. In contrast, fine-tuned BERT-based models, relying on high-quality in-domain data, show excellent performance but suffer from edit pattern overfitting. This paper proposes a novel dynamic mixture approach that effectively combines the probability distributions of small models and LLMs during the beam search decoding phase, achieving a balanced enhancement of precise corrections from small models and the fluency of LLMs. This approach also eliminates the need for fine-tuning LLMs, saving significant time and resources, and facilitating domain adaptation. Comprehensive experiments demonstrate that our mixture approach significantly boosts error correction capabilities, achieving state-of-the-art results across multiple datasets. Our code is available at https://github.com/zhqiao-nlp/MSLLM.

Figures

Figures reproduced from arXiv: 2506.06887 by the authors.

Figure 1
Figure 1. Overview of our approach. The correct sen [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Cases from rSIGHAN15 and LEMON-New test sets. The LLM used is Baichuan2(7B). 4.3 Evaluation Metrics Following the mainstream evaluation metrics for CSC tasks, we report the Precision (P), Recall (R), and F1 scores of the correction subtask at both the sentence- and character-level, denoted as S-P/R/F and C-P/R/F, respectively. To comprehensively evaluate the model’s correction capability, we also include the false p… view at source ↗
Figure 3
Figure 3. Model performance (sentence-level F1) of ReLM + LLMs on rSIGHAN15, ECSpell, and LEMON with different α and β. The x-axis is α and y-axis is β. The red cells denote superior performance of ReLM + LLM compared to ReLM. The blue cells represent inferior performance. The darker the color, the larger the performance gap. model in Eq. (1), which allows the language model to play a greater role [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: F1 scores of different LLMs (model size is 7B) with varying beam sizes. All models are combined with ReLM. Solid lines show sentence-level results, while dashed lines show character-level results. 7.2 LLM Approaches In the era of LLMs, researchers have been actively ex…
Figure 5
Figure 5. Figure 5: Prompt template. it introduces different synthesized errors for the same correct sentences that appear in both the train￾ing and test sets. These overlapping sentences ac￾count for 52.7%, 19.3%, and 28.2% of the ECSpell￾Law/Med/Odw training sets, respectively. For ex￾a…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  4. [4]

    Zuyi Bao, Chen Li, and Rui Wang. 2020. https://aclanthology.org/2020.findings-emnlp.184 Chunk-based Chinese Spelling Check with Global Optimization . In Findings of EMNLP, pages 2031--2040, Online

    work page 2020

  5. [5]

    Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, Xiaoyi Dong, Haodong Duan, Qi Fan, Zhaoye Fei, Yang Gao, Jiaye Ge, Chenya Gu, Yuzhe Gu, Tao Gui, Aijia Guo, Qipeng Guo, Conghui He, Yingfan Hu, Ting Huang, Tao Jiang, Penglong Jiao, Zhenjiang Jin, Zhikai Lei, Jiaxing Li, Jingwen Li, Linyang Li, S...

    arXiv 2024

  6. [6]

    Xingyi Cheng, Weidi Xu, Kunlong Chen, Shaohua Jiang, Feng Wang, Taifeng Wang, Wei Chu, and Yuan Qi. 2020. https://aclanthology.org/2020.acl-main.81 S pell GCN : Incorporating Phonological and Visual Similarities into Language Models for C hinese Spelling Check . In Proceedings of ACL, pages 871--881, Online

    work page 2020

  7. [7]

    Jacob Devlin, Ming - Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://aclanthology.org/N19-1423 BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding . In Proceedings of NAACL-HLT, pages 4171--4186, Minneapolis, Minnesota

    work page 2019

  8. [8]

    Ming Dong, Yujing Chen, Miao Zhang, Hao Sun, and Tingting He. 2024. https://doi.org/10.18653/v1/2024.findings-acl.439 Rich Semantic Knowledge Enhanced Large Language Models for Few-shot C hinese Spell Checking . In Findings of the Association for Computational Linguistics: ACL 2024, pages 7372--7383, Bangkok, Thailand. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.findings-acl.439 2024

Show all 34 references
  1. [9]

    Yong Hu, Fandong Meng, and Jie Zhou. 2024. https://aclanthology.org/2024.acl-long.10 CSCD - NS : a C hinese Spelling Check Dataset for Native Speakers . In Proceedings of ACL, pages 146--159, Bangkok, Thailand

    2024

  2. [10]

    Haojing Huang, Jingheng Ye, Qingyu Zhou, Yinghui Li, Yangning Li, Feng Zhou, and Hai-Tao Zheng. 2023. https://aclanthology.org/2023.findings-emnlp.771 A Frustratingly Easy Plug-and-Play Detection-and-Reasoning Module for C hinese Spelling Check . In Findings of EMNLP, pages 11...

    2023

  3. [11]

    Li Huang, Junjie Li, Weiwei Jiang, Zhiyu Zhang, Minchuan Chen, Shaojun Wang, and Jing Xiao. 2021. https://doi.org/10.18653/v1/2021.acl-long.464 PHMOS pell: Phonological and Morphological Knowledge Guided C hinese Spelling Check . In Proceedings of the 59th Annual Meeting of th...

    2021 doi

  4. [12]

    Wangjie Jiang, Zhihao Ye, Zijing Ou, Ruihui Zhao, Jianguang Zheng, Yi Liu, Bang Liu, Siheng Li, Yujiu Yang, and Yefeng Zheng. 2022. https://doi.org/10.1145/3511808.3557636 MCSCSet: A Specialist-annotated Dataset for Medical-domain C hinese Spelling Correction . In Proceedings ...

    2022

  5. [13]

    Jiahao Li, Quan Wang, Zhendong Mao, Junbo Guo, Yanyan Yang, and Yongdong Zhang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.287 Improving C hinese Spelling Check by Character Pronunciation Prediction: The Effects of Adaptivity and Granularity . In Proceedings of EMNLP, p...

    2022 doi

  6. [14]

    Kunting Li, Yong Hu, Liang He, Fandong Meng, and Jie Zhou. 2024. http://arxiv.org/abs/2406.16536 C-LLM: Learn to Check Chinese Spelling Errors Character by Character

    2024 arXiv

  7. [15]

    Yinghui Li, Haojing Huang, Shirong Ma, Yong Jiang, Yangning Li, Feng Zhou, Hai-Tao Zheng, and Qingyu Zhou. 2023. http://arxiv.org/abs/2307.09007 On the (In)Effectiveness of Large Language Models for Chinese Text Correction

    2023 arXiv

  8. [16]

    Zihong Liang, Xiaojun Quan, and Qifan Wang. 2023. https://doi.org/10.18653/v1/2023.acl-long.755 Disentangled Phonetic Representation for C hinese Spelling Correction . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2023 doi

  9. [17]

    Changchun Liu, Kai Zhang, Junzhe Jiang, Zirui Liu, Hanqing Tao, Min Gao, and Enhong Chen. 2024 a . https://doi.org/10.18653/v1/2024.emnlp-main.567 ARM : An Alignment-and-Replacement Module for C hinese Spelling Check Based on LLM s . In Proceedings of the 2024 Conference on Em...

    2024 doi

  10. [18]

    Linfeng Liu, Hongqiu Wu, and Hai Zhao. 2024 b . https://doi.org/10.1609/aaai.v38i17.29829 C hinese Spelling Correction as Rephrasing Language Model . In Proceedings of AAAI, pages 18662--18670, Vancouver, Canada

    2024 doi

  11. [19]

    Shulin Liu, Tao Yang, Tianchi Yue, Feng Zhang, and Di Wang. 2021. https://doi.org/10.18653/v1/2021.acl-long.233 PLOME: Pre-training with Misspelled Knowledge for C hinese Spelling Correction . In Proceedings of ACL-IJCNLP, pages 2991--3000, Online

    2021 doi

  12. [20]

    Qi Lv, Ziqiang Cao, Lei Geng, Chunhui Ai, Xu Yan, and Guohong Fu. 2023. https://doi.org/10.1145/3564271 General and Domain-adaptive C hinese Spelling Check with Error-consistent Pretraining . TALLIP, pages 1--18

    2023 doi

  13. [21]

    Ziheng Qiao, Houquan Zhou, Yumeng Liu, Zhenghua Li, Min Zhang, Bo Zhang, Chen Li, Ji Zhang, and Fei Huang. 2024. http://arxiv.org/abs/2412.12863 DISC: Plug-and-Play Decoding Intervention with Similarity of Characters for Chinese Spelling Check

    2024 arXiv

  14. [22]

    Yuen-Hsien Tseng, Lung-Hao Lee, Li-Ping Chang, and Hsin-Hsi Chen. 2015. https://aclanthology.org/W15-3106 Introduction to SIGHAN 2015 Bake-off for C hinese Spelling Check . In Proceedings of SIGHAN, pages 32--37, Beijing, China

    2015

  15. [23]

    Dingmin Wang, Yan Song, Jing Li, Jialong Han, and Haisong Zhang. 2018. https://aclanthology.org/D18-1273 A Hybrid Approach to Automatic Corpus Generation for C hinese Spelling Check . In Proceedings of EMNLP, pages 2517--2527, Brussels, Belgium

    2018

  16. [24]

    Dingmin Wang, Yi Tay, and Li Zhong. 2019. https://aclanthology.org/P19-1578 Confusionset-guided Pointer Networks for C hinese Spelling Check . In Proceedings of ACL, pages 5780--5785, Florence, Italy

    2019

  17. [25]

    Hongqiu Wu, Shaohua Zhang, Yuchen Zhang, and Hai Zhao. 2023. https://aclanthology.org/2023.acl-long.600 Rethinking Masked Language Modeling for C hinese Spelling Correction . In Proceedings of ACL, pages 10743--10756, Toronto, Canada

    2023

  18. [26]

    Shih-Hung Wu, Chao-Lin Liu, and Lung-Hao Lee. 2013. https://aclanthology.org/W13-4406 C hinese Spelling Check Evaluation at SIGHAN Bake-off 2013 . In Proceedings of SIGHAN, pages 35--42, Nagoya, Japan

    2013

  19. [27]

    Heng-Da Xu, Zhongli Li, Qingyu Zhou, Chao Li, Zizhen Wang, Yunbo Cao, Heyan Huang, and Xian-Ling Mao. 2021. https://aclanthology.org/2021.findings-acl.64 Read, Listen, and See: Leveraging Multimodal Information Helps C hinese Spell Checking . In Findings of ACL-IJCNLP, pages 7...

    2021

  20. [28]

    Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, Fan Yang, Fei Deng, Feng Wang, Feng Liu, Guangwei Ai, Guosheng Dong, Haizhou Zhao, Hang Xu, Haoze Sun, Hongda Zhang, Hui Liu, Jiaming Ji, Jian Xie, JunTao Dai, Kun Fa...

    2023 arXiv

  21. [29]

    Liner Yang, Xin Liu, Tianxin Liao, Zhenghao Liu, Mengyan Wang, Xuezhi Fang, and Erhong Yang. 2023 b . https://doi.org/10.1016/j.aiopen.2023.10.004 Is C hinese Spelling Check ready? Understanding the correction behavior in real-world scenarios . AI Open, pages 183--192

    2023 doi

  22. [30]

    Liang-Chih Yu, Lung-Hao Lee, Yuen-Hsien Tseng, and Hsin-Hsi Chen. 2014. https://aclanthology.org/W14-6820 Overview of SIGHAN 2014 Bake-off for C hinese Spelling Check . In Proceedings of CIPS - SIGHAN , pages 126--132, Wuhan, China

    2014

  23. [31]

    Shaohua Zhang, Haoran Huang, Jicong Liu, and Hang Li. 2020. https://aclanthology.org/2020.acl-main.82 Spelling Error Correction with Soft-Masked BERT . In Proceedings of ACL, pages 882--890, Online

    2020

  24. [32]

    Yue Zhang, Zhenghua Li, Zuyi Bao, Jiacheng Li, Bo Zhang, Chen Li, Fei Huang, and Min Zhang. 2022. https://aclanthology.org/2022.naacl-main.227 M u CGEC : a Multi-Reference Multi-Source Evaluation Dataset for C hinese Grammatical Error Correction . In Proceedings of NAACL, page...

    2022

  25. [33]

    Houquan Zhou, Zhenghua Li, Bo Zhang, Chen Li, Shaopeng Lai, Ji Zhang, Fei Huang, and Min Zhang. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.966 A simple yet effective training-free prompt-free approach to C hinese spelling correction based on large language models . In P...

    2024 doi

  26. [34]

    Houquan Zhou, Yumeng Liu, Zhenghua Li, Min Zhang, Bo Zhang, Chen Li, Ji Zhang, and Fei Huang. 2023. https://aclanthology.org/2023.findings-emnlp.495 Improving S eq2 S eq Grammatical Error Correction via Decoding Interventions . In Findings of EMNLP, pages 7393--7405, Singapore

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.