Chinese Academic Editing Station (chinese-academic-editing)

This built-in Chinese academic editing station (included since v0.10.0) is designed for use after peer-review/revise convergence and before ★final manual Chinese proofreading. It removes AI-generated traces from Chinese text and adjusts the register to formal academic written Chinese, while preserving the precise, evidence-bound tone that is correct in scholarly writing. Removing AI traces does not mean making it sound conversational—the voice of Chinese academic writing is formal, precise, and each claim is bound to evidence or markers; it is not colloquial, not personal.

Integration Contract [Source: academic-pipeline editing station contract]

This is an optional station: can be skipped, but cannot replace ★final manual Chinese proofreading.
Only linguistic presentation may be modified; do not alter numbers, // markers, [@citekey], claim strength, validation boundary, or method/limitation semantics.
Editing produces a new Chinese version manuscript/zh/vN.md (synchronizes new version if SI exists), without overwriting the old version.
After editing, rerun numbers, citations, claims, and marker checks, and a full check-consistency(zh); then ★final manual Chinese proofreading (state.zh_final) is still required before translate-en.
Any changes involving claim strength, numerical interpretation, methods, or limitations revert to the upstream source artifact (claim-audit/interpret-results/validation/analyze-data), and are not adjudicated in this station.
Terminology follows the glossary in project.yaml; if this station finds conflict between the glossary and conventional usage, it only marks it, without modifying the glossary itself.

Why no claim strength adjustment

Same reason as the English station: upgrading/downgrading verbs to "just right" changes claim strength, violating the "presentation only" contract. Moreover, by the time text reaches this station, claims have been reviewed in claim-audit and confirmed manually. If an over-claim or under-claim is detected, this station only marks it back to claim-audit/interpret-results, without rewriting. Strength adjudication belongs upstream; check_numbers.py/check_claims.py provide mechanical downstream checks.

When to use

Editing Chinese academic prose: sections of a manuscript, abstract, SI descriptions, Chinese versions of review responses. Not for blogs, marketing, personal essays, and it is not a tool for circumventing AI use disclosure.

Division of labor with `chinese-deai-editing`

chinese-deai-editing is an external user skill aimed at producing a reading feel of "human-edited" text, serving both Claude Code-generated technical .md files, with a bias towards "clear, direct, actionable." This station does not share that standard: the target reading feel is "like a journal paper." In case of conflict (e.g., "this sentence is too formal, shorten it" vs "this is a conventional academic expression"), the academic register always takes precedence. Non-manuscript Chinese materials (READMEs, design notes, external communications) should still use chinese-deai-editing.

Core Principle

Chinese academic writing already has a correct human voice: formal, precise, consistent terminology, each claim bound to evidence or markers. This station's job is (1) to remove Chinese AI tells without colloquialization or injecting personality; (2) to uphold the discipline that typical humanizers miss—preserve every number, marker, qualifier, and citation, and never let a verb exceed its evidence (editing only at the presentation layer; strength adjudication belongs to claim-audit).

Workflow

Read: Read the Chinese manuscript and, if available, the author's existing Chinese paper samples; read project.yaml for glossary, target, and analysis.rounding; note the target journal/discipline register.
Audit (do not edit yet): List each detected pattern (location + suggested fix). Items involving claim strength, numerical interpretation, methods, or limitations are marked "→ revert to upstream" and are not included in this station's rewrite list.
Rewrite: Structure and content unchanged; all numbers, markers, citekeys, and qualifiers are preserved; AI tells are removed, legitimate hedging is kept. Same number of sections, same number of paragraphs.
Report: Output the polished manuscript plus a change report (pattern types removed, items marked back to upstream, register notes), and confirm that no numbers, equations, markers, or citations have been modified.

Layer 1: General Chinese AI Tell List

Scan and correct the following patterns, subject to Layer 3's academic exceptions.

Empty openers: "In today's era of…", "With the rapid development of…", "In recent years, X has received widespread attention," "As is well known."

Meaning inflation: "of great significance," "crucial," "indispensable," "played a key role," "laid the foundation for…," "opened new directions."

Vague transitions and closings: "It is worth noting that," "It is worth mentioning that," "It is not hard to see that," "It can be seen from this that," "In conclusion," "To sum up," "The future can be expected."

Empty verb constructions: "carry out an analysis of X" → "analyze X"; "handle/treat accordingly" → directly use the verb; "made a contribution of…"

Translationese: "This causes…", "As a…", "is carried out," "based on this," "on the basis of…," "can be considered a kind of…"

Overused progression and parallelism: Everything is "not only…but also…"; the rule of three ("faster, more stable, more accurate"); mechanical antithesis ("A is the foundation, B is the core, C is the key"). Keep only if real progression exists; otherwise, use "also/and."

Consecutive connectives: Consecutive sentence starts with "In addition," "Moreover," "Meanwhile," "Furthermore"—let the logic carry itself.

Synonym cycling: Rotating different words for the same referent (model/architecture/framework used interchangeably for the same thing). Chinese academic writing requires consistent terminology; contrary to general writing, do not change terms just to avoid repetition.

Long sentence stacking: Strings of three or more clauses via "through…, and…, thereby…, then…" Split them. One idea per sentence.

Punctuation and formatting: Half-width punctuation marks (comma, period, exclamation, question mark) in Chinese text → full-width; quotation marks "…" → 「…」; overuse of em-dash insertions → changed to commas, colons, or split sentences; use dun-hao (、) for listed items, not comma.

Before: It is worth noting that the method proposed in this study not only significantly improved prediction accuracy, but also laid an important foundation for subsequent research in this field, and has broad application prospects. After: The prediction accuracy of this method is higher than existing methods ().

Layer 2: Chinese Academic AI Tells (Remove or Correct)

2.1 Overassertive verbs

Empirical research "shows," "supports," "is consistent with," not "proves," "confirms" universal conclusions. Watch: proves, confirms, establishes, fully demonstrates, completely validates; "significant" without statistical test or numerical value. (This station only adjusts tone at the presentation layer; if actual evidence does not support the verb, it is a claim strength issue, marked back to claim-audit, not rewritten here. See Layer 4 for strength alignment table.)

2.2 Meaning puffery

Watch: provides new ideas for…, fills a gap in…, has broad application prospects, opens a new path for…, has far-reaching impact, achieves a major breakthrough. Before: This study fills a gap in the field and provides new ideas for subsequent research. After: Previous work has not addressed online calibration under delayed annotation; this study addresses this scenario (Section 4).

2.3 Vague quantifiers and intensifiers

Watch: a large number of, a series of, many, numerous, extensive, comprehensive, fully. Before: We conducted a large number of experiments to comprehensively verify the effectiveness of the method. After: We evaluated on three datasets (). Note: When replacing with specific numbers, that number must already exist in results.json with a marker. If a key is missing, stop and mark back to analyze-data; do not convert or estimate on your own (per the hard rule from write-manuscript).

2.4 Novelty inflation

Watch: for the first time, innovatively, to the best of our knowledge, pioneering, unprecedented; multiple instances of "novel/innovative" in the same section.

2.5 Formulaic openers

Watch: "In recent years, the X field has received widespread attention," "With the continuous development of… technology," "Although there has been much research, however…" Replace by pointing directly to specific limitations of existing work.

2.6 Connective overuse

Do not let consecutive sentences start with "In addition/Moreover/Meanwhile/Furthermore."

2.7 Template contribution bullet points

Each contribution point must state a specific result, not a rephrased abstract. Before: The contributions of this study are as follows: (1) propose a new method; (2) conduct a large number of experiments; (3) achieve good results. After: This study (1) proposes an encoder conditioned on feature type; (2) shows smaller performance degradation compared to existing methods at 20% annotation noise (); (3) makes this benchmark dataset publicly available.

2.8 Citation stacking

Watch: "Many studies have shown [@a; @b; @c; @d]," "Research has pointed out" followed by a string of brackets. Rewrite to clarify which one or two studies are in conversation and why. Do not rewrite a sentence with citekeys into a general statement without source—this would lose both the lit_anchor and the citations (per write-manuscript step 5).

2.9 Vague hedging

Watch: to some extent, to a certain degree, relatively, somewhat, comparatively, one could say. Quantify or delete. Note: This differs from Layer 3's "evidence-bound hedging"—the distinction is whether it is tied to evidence. "Improved to some extent" is empty; "may reflect seasonal effects during the sampling period ()" is substantive and must be kept.

2.10 Template emphasis

Watch: It should be noted that, It is particularly worth emphasizing that, Importantly, It must be stated that. If it is truly important, the sentence itself will show it.

2.11 Long sentences and clause stacking

Split sentences exceeding approximately 60 characters or containing more than 3 clauses. Remove non-substantive modifiers, but do not remove qualifiers (see Layer 3).

2.12 Mainland Chinese terms → Taiwanese academic written usage

software, information, user, screen, default, code, memory, call, report error, quality (as in quality of work)→quality, robustness→robustness, confidence interval, significance level, sample size, fitting→goodness-of-fit, variance→variance, mean, standard deviation (retain).

Three exceptions, taking precedence over this list:

Glossary takes priority: Any term registered in project.yaml must be used as-is; when conflict arises with this list, follow the glossary, record the conflict in the change report, and do not modify the glossary.
Physical terms unchanged: "mass" is a correct technical term; do not change it to "quality." Similarly, "analysis of variance" if so registered in the glossary.
Commonly accepted terms kept: "data," "optimization," "interface" are already commonly used in Taiwanese academic/technical written Chinese; do not overcorrect.

Layer 3: Preserve, Do Not Overcorrect

Typical humanizers flatten legitimate academic structures; these must be kept:

Evidence-bound hedging is correct and necessary. When a claim is genuinely uncertain, retain "shows," "supports," "is consistent with," "may reflect," "this study cannot rule out," "awaits further testing." Incorrect fix: Changing "Results show X" to "Results prove X"—that creates an over-claim.
Qualifiers and boundary sentences are substantive content. Boundaries of qualified validation, population/scope restrictions, and exploratory designations must not be deleted or compressed simply because they "read as verbose."
Passive voice is fine when the agent is irrelevant: "Samples were normalized to total protein."
"This study" and "We" are standard in Chinese academic writing; do not rewrite just to avoid them.
Methods section conventional expressions ("evaluated using…", "using the… model", "stratified sampling by…") are conventions, not AI tells.
Term repetition is better than synonym substitution (see Layer 1 "synonym cycling").
Formal definitions, named methods/metrics, equations, and symbols are kept exactly as-is; any first-occurrence English original paired with Chinese translation is retained.
Never invent, delete, or change any number, equation, or citation; retain every marker and citekey. (Downstream mechanical checks via check_numbers.py/check_citations.py/check_claims.py and full check-consistency(zh) provide downstream mechanical safeguarding.)

Layer 4: Results and Discussion Statement Specifications [New to this station]

This layer codifies "stating results clearly in academic tone" into an enforceable presentation-layer specification. It does not change claims, only specifies which sentence pattern to use for a given claim. All rules anchor to write-manuscript steps 4/5/6 and the existing contract of check_numbers.py; no new contracts are created.

4.1 Results section sentence patterns

One finding per sentence, pattern: Subject + direction + value (with marker) + uncertainty + source.

Do not use bare p-value as the main clause; put p-value in parentheses as supplementary. Before: Analysis results showed a significant difference between the two groups (p < 0.01), indicating that the treatment was effective. After: The X of the treatment group was higher than that of the control group (, 95% confidence interval ; p ).
Effect size and uncertainty appear together. If results.json lacks keys for effect size or uncertainty, stop, mark back to analyze-data; do not convert, estimate, or substitute with "significant."
Numerical precision unchanged: if a value has a display key, it must be exactly that string; otherwise, follow analysis.rounding. Editing must not change precision, e.g., changing "0.043" to "about 0.04."
Results section does not explain causation or mechanism—that is for the Discussion section. Remove phrases like "this shows," "this proves," "this means" from the Results section.
Table blocks (scope of ) are not rewritten cell by cell; only edit guiding sentences outside the table.
Figure captions are taken from the figure manifest's caption field; not rewritten at this station.

4.2 Discussion section sentence patterns

Each scientific claim ends with ; wording must not exceed the strength of that claim.
Limitations section must not be turned into positive framing. Common AI bias is to rewrite flaws as selling points ("can be further extended to larger samples" instead of "the sample size of this study is insufficient for subgroup comparisons"). Limitations must remain limitations.
The "conversation with literature" paragraph must not rewrite lit_anchor citations into unsourced general statements; counterevidence paragraphs must not be weakened.
"Future directions" must not be written as commitments or application prospect pitches.

4.3 Abstract

Word count follows target.abstract_limit; numbers, claim strength, and population/scope restrictions are consistent with the main text. Do not remove core restrictions retained in the main text from the abstract—this is the most common editing casualty.

4.4 Strength alignment table (presentation-layer wording)

Level	Chinese verb/phrase	Usage
Strong	proves, confirms, establishes	General empirical studies do not use; except for mathematical proof-type studies
Medium-high	shows, indicates, supports	Direct test or clear effect size exists
Medium	is consistent with, suggests, tends to	Compatibility evidence, same direction but insufficient power
Low	may reflect, awaits further testing, cannot rule out	Exploratory, alternative explanations not ruled out

Rule: Substitution within the same level is permitted for stylistic reasons; cross-level changes are never allowed. If the original wording does not match claim strength, mark back to claim-audit—do not upgrade or downgrade on your own.

Layer 5: Register and Discipline Alignment

If the author provides an existing Chinese paper, read the sample first, noting sentence length rhythm, connective habits, degree and placement of hedging, how sections begin, conventional notation, and then align. Also align with disciplinary register (experimental science: tight, results-forward; humanities/social sciences: more expository, allowing longer argumentative paragraphs). With no sample, default to clean, precise, formal academic Chinese, not the colloquial, opinionated voice of external humanizers.

Self-Check (Must do after editing)

Residual AI tells: Read again and ask, "Does any part still read as if it were generated?"—any unremoved clichés, overly clean parallelism, remaining empty positive words, missed half-width punctuation.
Problems introduced by the reviser: The reviser is also an LLM. Check item by item—has one cliché been swapped for another cliché? Are all sentence lengths roughly the same? Has hedging been flattened into assertions, qualifiers deleted as verbose, or markers/citekeys changed along with the sentence? Point 2 is this station's most severe failure mode; check it first.

Output

Output the polished Chinese manuscript (new version manuscript/zh/vN.md) along with a brief change report:

Patterns removed (category statistics).
Claim-related but unchanged items marked back to claim-audit/interpret-results (list each with location).
Items where missing results.json keys prevented implementing §4.1 sentence patterns, marked back to analyze-data.
Terms conflicting with glossary (list each with suggestion and rationale; do not modify glossary).
Register notes (which sample/discipline convention was aligned to).
Confirmation that no numbers, equations, markers, or citations were modified.

Provenance

Integration contract, core principle, workflow, Layer 1/2/3/5 skeleton: Symmetrically derived from this plugin's built-in academic-english-editing (its Layers 1/2/3/5 originated from user skill academic-humanizer v0.3.3, MIT); Chinese context patterns, example sentences, and rules are newly written.
Chinese AI tell list and Taiwanese written Chinese normalization: Partially references external user skill chinese-deai-editing (same author), but the register baseline has been changed from "human-edited feel" to "journal paper feel," and its technical .md register has been removed.
Layer 4 (Results and Discussion statement specifications): New contribution of this station [Source: academic-pipeline editing station contract]. Rules anchor to write-manuscript steps 4/5/6 and existing waiver definitions in check_numbers.py, no new contracts created; this layer is updated synchronously with write-manuscript revisions.
Intentionally excluded: Claim strength adjustment (same reason as the English station excluding academic-humanizer Layer 4); drafting mode (this station only polishes existing manuscripts; drafting remains with write-manuscript); technical document register (delegated to external chinese-deai-editing).

Ask in your favorite AI

Documentation