Chinese text correction has traditionally focused on spelling and grammar, while factual error correction is usually treated separately. However, in paragraph-level Chinese professional writing, linguistic (word/grammar/punctuation) and factual errors frequently co-occur and interact, while many draft-level errors are sparsely observable in published texts after editorial review, making unified correction both necessary and controlled benchmark construction essential. This paper introduces CLFEC (Chinese Linguistic & Factual Error Correction), a new task for joint linguistic and factual correction. We construct a mixed, multi-domain Chinese professional writing dataset spanning current affairs, finance, law, and medicine. We then conduct a systematic study of LLM-based correction paradigms, from prompting to retrieval-augmented generation (RAG) and agentic workflows. The analysis reveals practical challenges, including limited generalization of specialized correction models, the need for evidence grounding for factual repair, the difficulty of mixed-error paragraphs, and over-correction on clean inputs. Results further show that handling linguistic and factual errors within the same context outperforms decoupled pipelines, and that agentic workflows can be effective with suitable backbone models. Overall, CLFEC provides a new benchmark for Chinese text correction research and practical guidance for proofreading systems.
Figures & tables
Error type
Example snippet
Word
反溃 控制 → 反馈控制 collapse control → feedback control
Grammatical
增强 …水平 → 提高…水平 Enhance …level → Raise…level
Punctuation
如下 ； → 如下： As follows ; → As follows:
Factual
六中 全会 → 四中全会 The Sixth Plenary Session → The Fourth Plenary Session
Table 1: Unifying linguistic and factual error correction, where linguistic errors cover word, grammatical, and punctuation types.
Figure 1: Verified error density (per 10K characters) by source type and proofreading product. Professional texts yield few confirmed errors overall, while UGC texts show higher density.
Figure 2: Overview of the CLFEC data construction and annotation process. We collect clean professional texts from four domains, inject factual and linguistic errors under controlled settings, iteratively verify candidate edits, and convert span-level annotations into CHERRANT-style normalized edit pairs for evaluation.
Figure 3: Statistics of the CLFEC dataset.
Dataset
Ann.
Natural
Synthetic
Agreement
CLFEC
A1
143 (71.5%)
57 (28.5%)
84.0%
A2
143 (71.5%)
57 (28.5%)
RWS
A1
179 (89.5%)
21 (10.5%)
80.0%
A2
159 (79.5%)
41 (20.5%)
Table 2: Human evaluation of error naturalness. Agreement denotes the percentage of samples on which the two annotators assign the same label.
model
method
Detection
Correction
Correction (MIX)
Correction (LEC)
Correction (FEC)
P
R
F1
P
R
F1
P
R
F1
P
R
F1
P
R
F1
LLM-based systems
CEC3-4B
prompt
34.06
14.90
20.73
31.78
14.04
19.47
39.61
13.62
20.27
38.95
21.58
27.77
–
–
–
Qwen3-4B
prompt
37.74
23.15
28.70
31.71
20.20
24.68
43.52
18.93
26.39
37.74
30.43
33.69
6.97
7.14
7.05
S-RAG
35.72
47.01
40.60
31.40
43.81
36.58
40.07
42.92
41.45
32.32
44.13
37.32
19.89
47.86
28.11
U-RAG
43.17
39.85
41.44
39.16
37.54
38.33
49.68
35.63
41.50
39.33
39.75
39.54
28.07
44.02
34.28
Table 3: Overall results on CLFEC . We report Precision (P), Recall (R), and F1 for detection and correction, and further break down correction performance on three subsets: MIX , LEC , and FEC . Best and second-best numbers are highlighted in bold and underline, respectively.
Figure 4: Detection recall by error type on CLFEC.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Method
Word
Factual
Punctuation
Grammatical
LLM-based systems
CEC3-4B
prompt
32.94
0.00
2.30
8.45
Qwen3-4B
prompt
41.00
5.09
22.62
16.43
S-RAG
59.00
44.77
32.79
27.70
U-RAG
47.51
41.96
24.92
23.47
DS-V3.2
prompt
70.02
19.44
59.02
47.42
Appendix
Table 4: Recall (%) for each error type across models and strategies. Best in each column is bold; second best is underlined.
Model
Method
MIX
LEC
FEC
Error-free
Gold
gold
0.72
0.47
0.33
0.00
CEC3-4B
prompt
0.48
0.52
0.43
0.34
Qwen3-4B
prompt
0.24
0.26
0.13
0.23
S-RAG
0.72
0.61
0.77
0.32
U-RAG
0.50
0.46
0.51
0.25
DS-V3.2
prompt
0.68
0.62
0.61
0.46
Appendix
Table 5: The number of edits performed per 100 characters on MIX/LEC/FEC and an Error-free subset. Higher values on Error-free subset indicate stronger over-correction.
Subtype
Weight
Homophonic confusable (同音字词)
0.15
Shape-similar confusable (形近字词)
0.15
Near-phonetic confusable (近音字词)
0.10
Redundancy (字词冗余)
0.15
Omission (字词缺失)
0.15
Permutation (字词乱序)
0.05
Appendix
Table 6: Word/punctuation error subtypes and sampling weights used in our injection prompts.
Figure 5: Word & Punctuation Error Injection Prompt
Figure 6: Word & Punctuation Error Injection Prompt (English Translation)
School of Computer Science and Technology, East China Normal University · Shanghai Institute of Artificial Intelligence for Education, East China Normal University