Chinese Traditional Conversion

Convert to Traditional Chinese

Convert simplified Chinese into Traditional Chinese as written in Taiwan. The script settles what is deterministic and reports what is not; deciding the reported items is your job.

Never let the script guess and never guess yourself. A character like 发 is 發 or 髮 depending on the sentence around it. Tools that pick one silently are why this skill exists.

Division of labour

scripts/convert.py owns the mechanical half: 6300+ deterministic character mappings, Taiwan variant folding (着→著, 裏→裡, 爲→為), unambiguous vocabulary, and exhaustive detection. It is exhaustive where you are not — it will not miss the one stray 发 on line 400.

You own the contextual half: every ambiguous character, every vocabulary item with more than one Taiwan reading, and the phrasing that no dictionary encodes. You have the whole document; the script has a table.

Workflow

1. Establish the input

Files or directories go through the path form. Text pasted into the conversation goes through --stdin. Convert files in place so the user reviews a git diff; never paste a converted copy of a file back into the chat.

2. Run the mechanical pass

scripts/convert.py PATH...          # convert in place, print the review report
scripts/convert.py --check PATH...  # report only, write nothing
scripts/convert.py --diff PATH...   # unified diff, write nothing
scripts/convert.py --json PATH...   # structured findings, for triage without parsing prose
echo "文字" | scripts/convert.py --stdin   # result on stdout, report on stderr

On a directory of any size, run --check first. It tells you the scale of the job and which files are affected before anything is written.

Add --markdown to apply markdown protection rules to --stdin input. Add --include-code to convert protected regions too, which is almost never right.

Add --no-naer to suppress the terminology advisories described below, or --naer-all to widen them to text that was already traditional.

Exit codes: 0 clean, 1 review items only, 2 must-fix items present, 3 usage error. NAER advisories never change the exit code. The script checks its own prerequisites and explains failures on stderr, so do not pre-verify anything.

3. Resolve the report

Three sections, three different obligations.

MUST FIX — a simplified character no substitution resolves on its own. The document is not zh-TW until every one is decided. Read the rule the report prints, look at the sentence, edit the file.

REVIEW — the script applied a form that is valid zh-TW, which may still be the wrong one here. 面 stays 面 unless it is 麵條. Confirm or correct each one.

Occurrences a curated keep-rule already answers are detected but collapsed into a ✓ summary line, because a report that lists 同 four hundred times cannot honestly be worked through. The detection behind them is unchanged; --review-all lists them individually.

PROTECTED — simplified text inside code blocks, inline code, or URLs. Left untouched on purpose. A simplified string literal is often deliberate. Change these only when the user asks.

NAER — 國家教育研究院 has ruled on this term and its Taiwan rendering differs from what the text uses. The advisory prints the official rendering, the English headword, and the subject domain. Nothing was substituted, and nothing has to be: this corpus records the preferred academic rendering, not a correction. 除臭 is not wrong because a glossary prefers 去臭. Apply one when the domain matches the document and the rendering is clearly better — 纳米 really should be 奈米 — and ignore it otherwise.

Advisories are raised only where the span actually held a character that cannot stand in zh-TW. These glossaries are specialist and their mainland column collides constantly with ordinary words: the electrical-engineering set lists 日期 as data, a typo for 数据, which would otherwise flag every date in a document. Text already written in Traditional needs no localising. This is narrower than the other tiers on purpose — vocabulary that matters for ordinary writing lives in term_map.json and is converted regardless of how it is spelt, so 軟件 still becomes 軟體 even with no simplified character in sight. Pass --naer-all when reviewing Hong Kong or mainland material that is already written in Traditional.

An entry with no rule printed is one the character table has not been curated for yet. Decide it from context as usual, and prefer the form used elsewhere in the same document.

樂詞網 settles what the report cannot. For a terminology question the tables do not answer, query https://terms.naer.edu.tw/search/?q=TERM with WebFetch — it is the authority for Taiwan academic and technical vocabulary. It is HTML with no API, so read the page rather than expecting JSON. It does not rule on character-level ambiguity: whether 发 is 發 or 髮 is a question for 教育部《重編國語辭典修訂本》, not for a terminology database.

4. Localise the phrasing

Character and vocabulary conversion produces correct but foreign-sounding Chinese. Read references/taiwan-vocabulary.md and apply the sentence-level adjustments: 通过 as a preposition becomes 透過 or 藉由, 对…进行… constructions unwind into direct verbs.

5. Verify

Re-run --check. Zero must-fix items is the gate. Review items may legitimately remain when you have consciously judged each one, so exit 1 is an acceptable end state and exit 2 is not.

6. Report to the user

Say what was converted, which items you judged and why, and what you deliberately left alone. Name any proper noun you were unsure about.

Rules

• The document's own usage wins. Before accepting a vocabulary change, check what the author already does: grep -c 函式 FILE against grep -c 函數 FILE. A document that consistently says 基類 has a convention, and imposing 基底類別 on it is rewriting, not converting.
• Conversion is not rewriting. Preserve the author's voice, structure, and argument. Do not apply writing-style guidance. When the user wants the prose improved as well, that is chinese-content-writing-guideline, invoked separately and said out loud.
• Proper nouns are reported, never converted. 沈 is the surname 沈 or the city 瀋陽; 姜 is the surname 姜 or the spice 薑. The report flags these with a proper-noun warning. If the text does not settle it, ask the user.
• Code, URLs, filenames, and identifiers stay simplified. Converting a dictionary key or a path breaks the program.
• Quoted material gets characters only. Statutes, judgments, and other people's words take the character conversion and nothing else. Tell the user you limited it.
• Every decision needs a reason you can state. If you cannot state one, leave it and flag it for the user rather than guessing.

Data tables

assets/ holds three tables the script reads. They are data, not reading material — do not load them into context.

scripts/build_tables.py regenerates the character tables from the Unicode Unihan database. It is maintenance tooling, needs network access, and is never part of a conversion. Run it only to refresh the tables against a new Unicode release; it preserves every human decision already recorded. --lint-terms checks term_map.json for entries that contradict each other.

--naer rebuilds naer_terms.json from 國家教育研究院's cross-strait terminology on data.gov.tw. It verifies each dataset carries licence 1 (政府資料開放授權條款第1版) before using it, and drops any mainland form that is also a Taiwan rendering elsewhere in the corpus, because no table can tell 訪問 the noun from 訪問 the verb. Add --naer-rescan to re-discover the dataset list. A download that fails aborts the rebuild and leaves the existing table in place, so a flaky connection cannot quietly ship a shrunken one; --naer-allow-partial overrides that and records partial: true in the file.

Conversion itself never touches the network. Only build_tables.py does, and it treats the upstreams as someone else's infrastructure: responses are revalidated against a disk cache in ~/.cache/chinese-traditional-conversion, so a rebuild that changes nothing transfers nothing instead of re-fetching 15 MB; retries back off rather than hammering a struggling host, and a permanent answer such as 404 is not retried at all. The dataset scan runs at four concurrent requests and its result is cached, so it should rarely need re-running.

資料來源：國家教育研究院，依政府資料開放授權條款第1版釋出。