i18n: Simplified Chinese (zh-CN) adaptation as a separate repo #203

New issue

Closed

opened 2026-08-01 14:46:48 +00:00 by jiji262 · 0 comments

jiji262 commented

2026-08-01 14:46:48 +00:00

(Migrated from github.com)

Hi @blader — following #163 (French), #138 (Spanish) and #194 (Traditional Chinese), I built a Simplified Chinese (zh-CN) localization of humanizer. Flagging it here for discoverability, not to request a merge.

Repo: https://github.com/jiji262/humanizer-chinese

You noted in #163 that you'd rather localized variants stay as separate community repositories so each language can evolve without adding duplicate runtime authorities to the core project. This one fits that shape: standalone repo, MIT, with blader/humanizer attribution kept in the LICENSE, README, and SKILL.md.

Why it is an adaptation, not a translation

Three of your rules invert or soften in Simplified Chinese, so a literal port would actively damage correct prose:

Curly quotes (§19) invert. “” are the standard in Simplified Chinese (GB/T 15834-2011), not a tell. The rule is replaced with detection of full-width/half-width punctuation mixing, which is the real machine fingerprint.
The em dash ban (§14) is softened, not copied. The Chinese dash —— is legitimate punctuation for explanation and topic shifts. Only overuse, paired parenthetical dashes, and mis-typesetting (single-width —, or - / -- standing in for a dash) are flagged.
Subjectless sentences (§13) are idiomatic Chinese, not hidden-actor passives. Flagging them would gut normal writing, so that slot instead targets Europeanized 被-constructions and translationese.
Hyphenated word pairs (§26) are dropped — Chinese has no hyphenated compounds.

Net: 24 patterns map one-to-one, 8 collapse into 4, 1 is dropped, and 8 Chinese-specific patterns are added, for 36 total.

Chinese-specific patterns, with sources

Lecture-note signposting (首先…其次…最后, 值得注意的是, 综上所述) — AI conjunction density measured at roughly 3× human in a 6,586-text parallel corpus study (CCL 2023, Beijing Language and Culture University).
Parallelism and antithesis stacking — measured at roughly 6× human density in Chinese AI text, the strongest single quantitative tell found.
Bureaucratic empty verbs (对…进行分析, 作出调整) — the Europeanized-Chinese problem Yu Kwang-chung criticized in 1987, which LLM training data amplifies.
Uniformly written register — AI modal-particle density is about 1/5 of human in the same corpus study.

Two additions that may be of general interest

Both came out of baseline testing (running "de-AI-ify this" on Chinese AI text without a skill, then reading what went wrong), so they may generalize beyond Chinese:

A register gate. For encyclopedic, technical, legal, and reference text, neutral plain prose is the correct human voice. Without this, "humanizing" turns a reference article into a personal blog post.
An anti-overcorrection list. The most common failure was not leftover AI tone — it was inventing first-person anecdotes ("back when I was learning to code…") to manufacture human texture. That is fabrication wearing a human mask, so your no-fabrication rule from v2.9.0 is extended to name it explicitly. A second failure was flattening every genre into the same fake-colloquial voice, which is just a different template.

Distinct from #194 (zh-TW): different script, different punctuation standard (“” vs 「」), and different vocabulary; the two are siblings rather than duplicates.

Happy to close this if you'd rather keep the tracker clear — it's here mainly so Simplified Chinese users can find a version that doesn't break their punctuation.

Hi @blader — following #163 (French), #138 (Spanish) and #194 (Traditional Chinese), I built a **Simplified Chinese (zh-CN)** localization of humanizer. Flagging it here for discoverability, not to request a merge. Repo: https://github.com/jiji262/humanizer-chinese You noted in #163 that you'd rather localized variants stay as separate community repositories so each language can evolve without adding duplicate runtime authorities to the core project. This one fits that shape: standalone repo, MIT, with blader/humanizer attribution kept in the LICENSE, README, and SKILL.md. ### Why it is an adaptation, not a translation Three of your rules invert or soften in Simplified Chinese, so a literal port would actively damage correct prose: - **Curly quotes (§19) invert.** `“”` are the *standard* in Simplified Chinese (GB/T 15834-2011), not a tell. The rule is replaced with detection of full-width/half-width punctuation mixing, which is the real machine fingerprint. - **The em dash ban (§14) is softened, not copied.** The Chinese dash `——` is legitimate punctuation for explanation and topic shifts. Only overuse, paired parenthetical dashes, and mis-typesetting (single-width `—`, or `-` / `--` standing in for a dash) are flagged. - **Subjectless sentences (§13) are idiomatic Chinese,** not hidden-actor passives. Flagging them would gut normal writing, so that slot instead targets Europeanized `被`-constructions and translationese. - **Hyphenated word pairs (§26) are dropped** — Chinese has no hyphenated compounds. Net: 24 patterns map one-to-one, 8 collapse into 4, 1 is dropped, and 8 Chinese-specific patterns are added, for 36 total. ### Chinese-specific patterns, with sources - **Lecture-note signposting** (`首先…其次…最后`, `值得注意的是`, `综上所述`) — AI conjunction density measured at roughly 3× human in a 6,586-text parallel corpus study (CCL 2023, Beijing Language and Culture University). - **Parallelism and antithesis stacking** — measured at roughly 6× human density in Chinese AI text, the strongest single quantitative tell found. - **Bureaucratic empty verbs** (`对…进行分析`, `作出调整`) — the Europeanized-Chinese problem Yu Kwang-chung criticized in 1987, which LLM training data amplifies. - **Uniformly written register** — AI modal-particle density is about 1/5 of human in the same corpus study. ### Two additions that may be of general interest Both came out of baseline testing (running "de-AI-ify this" on Chinese AI text *without* a skill, then reading what went wrong), so they may generalize beyond Chinese: 1. **A register gate.** For encyclopedic, technical, legal, and reference text, neutral plain prose *is* the correct human voice. Without this, "humanizing" turns a reference article into a personal blog post. 2. **An anti-overcorrection list.** The most common failure was not leftover AI tone — it was **inventing first-person anecdotes** ("back when I was learning to code…") to manufacture human texture. That is fabrication wearing a human mask, so your no-fabrication rule from v2.9.0 is extended to name it explicitly. A second failure was flattening every genre into the same fake-colloquial voice, which is just a different template. Distinct from #194 (zh-TW): different script, different punctuation standard (`“”` vs `「」`), and different vocabulary; the two are siblings rather than duplicates. Happy to close this if you'd rather keep the tracker clear — it's here mainly so Simplified Chinese users can find a version that doesn't break their punctuation.