taiwan-corpora/twfilter-tables

twfilter reference tables

The tables that decide whether a span of traditional-Chinese text is Taiwan Mandarin (臺灣華語) rather than Hong Kong Cantonese, mainland text converted to traditional characters, or literary Chinese. Plain text, one record per line, tab-separated where a record has fields, LC_ALL=C sort order, UTF-8, LF.

Consumed by twfilter 0.1.0, where this directory is vendored byte-for-byte and MANIFEST.json is verified by its test suite. Usable standalone: nothing here depends on the gem.

Files

file	rows	contents
moe_common.txt	4 808	教育部 常用國字標準字體表
moe_secondary.txt	6 343	教育部 次常用國字標準字體表
moe_rare.txt	18 356	教育部 罕用字體表
moe_exception.txt	12	dictionary headword characters outside all three charts
simplified_only.txt	3 785	characters that exist only as simplified forms
variants_used_in_taiwan.txt	16	simplified-looking forms that are standard in Taiwan
converted_orthography.txt	14	mainland traditional orthography: 裏着衞爲説麽綫墻眞啓衆産囯麪
cantonese.txt	17	characters of written Cantonese
regional_hard.tsv	19	Hong Kong, Cantonese-sequence and Singapore forms
mainland_hard.tsv	149	mainland form, Taiwan form, attestations of each
mainland_soft.tsv	38	ambiguous pairs: mark, do not reject
mainland_exceptions.tsv	6	superstrings in which a marker is legitimate Taiwanese
foreign_topics.txt	80	non-Taiwanese toponyms and institutions — a topic tag, not an origin marker
taiwan_markers.txt	66	ROC institutions and Taiwanese place names
taiwan_lexicon.txt	51	Taiwan forms of Taiwan/mainland lexical oppositions
taiwan_particles.txt	7	sentence-final particles in their Taiwan spelling
taiwan_grammar.txt	20	Southern Min substrate constructions
erhua_headed.txt	14	兒 as an initial root morpheme (兒童, 兒歌)
erhua_tailed.txt	36	兒 as a final root morpheme (女兒, 那兒)
wenyan.txt	4	literary-Chinese particles for density estimation
MANIFEST.json	—	row count and SHA-256 per table

Membership

A table is only as good as the rule that admits a row to it. Each rule below is enforced against the shipped data by spec/twfilter/membership_spec.rb in the gem, so a table cannot drift away from the rule that justifies it.

table	a row is admitted iff
cantonese.txt	the character is used in written Cantonese and is not in 常用國字標準字體表
converted_orthography.txt	the form is in no MOE chart and the Taiwan form outnumbers it ≥ 20 : 1
variants_used_in_taiwan.txt	the form looks simplified but appears in an MOE chart
mainland_hard.tsv	the mainland form is unattested in Taiwan and the Taiwan form occurs ≥ 3 times
mainland_soft.tsv	the pair is a real opposition but the mainland form is attested in Taiwan
mainland_exceptions.tsv	the head term is a detector in one of the two bands
regional_hard.tsv	the form is unattested in Taiwan and its Taiwan counterpart occurs ≥ 3 times
taiwan_lexicon.txt	a mainland writer would not use this string for this referent
taiwan_markers.txt	the name denotes an ROC institution or a Taiwanese place
taiwan_particles.txt	the graph is the Taiwan-preferred spelling of the particle
taiwan_grammar.txt	the construction is a Southern Min substrate feature

The rules exclude far more than they admit, and the exclusions are the point:

• Subject matter is not provenance. Taiwanese food names are not evidence of Taiwanese authorship; neither are mainland place names evidence of mainland authorship.
• Pan-Mandarin forms are not evidence. A-not-A interrogatives (好不好, 對不對, 有沒有) are a general Mandarin question type, not a Taiwan feature. 高鐵, 郵局, 晚安, 番茄 and 夜市 are written identically on both sides of the strait.
• A character in everyday Taiwanese use cannot mark another variety, however characteristic of that variety it may be. 俾 and 唔 are 常用國字; 俾 is formulaic in ROC official prose (俾利, 俾便).
• A homograph is not a marker. 沙特 is Taiwan's rendering of Sartre as well as the mainland's rendering of Saudi Arabia; 重載 is standard in Taiwanese engineering.

Derivation

simplified_only.txt is the OpenCC STCharacters key set minus everything reachable in TSCharacters, minus the 16 forms in variants_used_in_taiwan.txt (秘 群 床 峰 粽 庄 晒 痴 霉 虱 灶 昵 肴 羡 洒 痒), each of which appears in an MOE chart despite occurring in the simplification tables.

The reference corpus

Mainland pairs and orthographic forms are graded by attestation against ten natively Taiwanese sources — statute, parliamentary record, municipal press, forum prose, literary prose, news, pedagogical material, dictionary headwords, read speech and vernacular literature — totalling 11 923 110 characters.

The corpus admits no source with an open authorship gate. Web comments on Taiwanese channels are Taiwan-hosted but not Taiwan-authored: in a 15 M-character superset, every mainland form attested anywhere occurs in open comments and nowhere else — 芯片, 幾率, 硬件, 英偉達, 帶寬, 光驅, 冰棍. Counting them would license the vocabulary of mainland commenters as Taiwanese. Hosting is not authorship.

Grading bands

• hard — the mainland form has zero occurrences and the Taiwan form at least three. Rejects.
• soft — the pair is a genuine opposition but the mainland form is attested here, or the two are hand-graded ambiguous. Marks; never rejects by default.
• excluded — everything else, including pairs where neither form is attested.

593 of 780 candidate pairs are excluded, and the excluded set is the point of the exercise. The OpenCC TWPhrases table is an orthographic conversion list built for rendering, not a provenance signal: used naively as an origin detector it flags 8.83 % of the ROC statute book, because 程序, 支持, 設備, 文件 and 質量 are ordinary Taiwanese words.

Known limitation

The reference corpus is itself the output of an earlier origin filter. The grading is therefore valid for pruning candidates that filter never saw, and invalid for confirming forms it had already removed. Forms that fail the ≥ 20 : 1 threshold are removed on the strength of positive attestation, which the circularity does not affect; forms retained at zero attestation carry the weaker warrant.

Mainland topics are not origin markers

Measured on 63 522 documents hosted on .tw domains — Taiwan-gated by DNS, never touched by any lexical filter — 北京 occurs in 2 722 documents across 693 hosts, 上海 in 923, 戶口 in 177, 高考 in 309. Each is a subject, not a provenance signal. Several are ordinary ROC administrative vocabulary: 戶口名簿 is a household registration booklet, 高考 is 公務人員高等考試, 派出所 is a police station. And Taipei streets are named after mainland cities by KMT-era decree, so 南京東路, 重慶南路, 南昌路 and 貴陽街 are Taipei addresses.

foreign_topics.txt is therefore a topic tag. 戶口 and 高考 were dropped outright, and a hit is suppressed when the term is followed by an address suffix.

Measured behaviour

Rejection rate over 987 740 sentences from fifteen sources, after punctuation normalisation, under the gem's corpus policy. orthography counts sentences rejected for mainland traditional forms.

source	composition	sentences	rejected	lexical	orthography
全國法規資料庫	native	102 090	0.12 %	0.00 %	0
立法院公報	native	120 569	0.63 %	0.00 %	0
教育部國語辭典簡編本	native	38 826	1.43 %	0.00 %	0
新北市政府新聞稿	native	33 870	1.96 %	0.00 %	11
PTT 八卦板	native	49 775	0.55 %	0.00 %	23
新聞語料	native	41 000	2.95 %	0.00 %	17
方格子散文	native	16 124	5.87 %	0.00 %	5
開放式網路留言	open authorship	156 105	7.05 %	0.01 %	528
字幕語料	translated	300 000	0.92 %	0.01 %	359
TED 講稿	translated	69 000	1.88 %	0.07 %	182
中文維基百科 (臺灣正體)	converted	24 527	4.92 %	0.08 %	180
all fifteen	—	987 740	2.36 %	0.01 %	—

Two results matter. The lexical detectors fire on no natively composed source and only on the sources composed outside Taiwan or written by uncontrolled authors. And mainland orthography now fires 67 times across 424 024 natively composed sentences against 1 278 times across 563 716 sentences that are not — a 14 × gradient. In the previous release the same check rejected 61 sentences of the statute book and 335 of the news corpus, on 覈, 峯 and 牀: 覈 is 次常用國字 and formulaic in ROC administrative prose (檢覈, 覈實), 峯 is a Taiwanese given-name variant, and 牀 is standard in 臨牀 and in 礦牀, which is statutory vocabulary.

The residual rejections on native sources are dominated by shape, not provenance: sentence fragments, list punctuation and read-speech prompts below the six-character floor.

What a verdict does and does not guarantee

A passing verdict on a single sentence means no evidence of non-Taiwanese origin was found. It is not a positive certificate of Taiwanese origin, and it cannot be one: 85–90 % of genuinely Taiwanese sentences carry no Taiwan-specific marker at all (89.6 % of the statute book, 84.4 % of the news corpus). A sentence of neutral prose is compatible with every written standard, and requiring a marker in each one would discard most of the language.

Measured against text these tables are built to exclude:

input	single sentences kept, corpus	with foreign_topics_reject	as a block
Hong Kong 書面語, no Cantonese particles	7 / 8	4 / 8	0 / 24
Singapore and Malaysia, traditional	2 / 4	1 / 4	—
mainland converted to traditional, formal register	2 / 4	1 / 4	—
Taiwanese control	2 / 2	2 / 2	24 / 24

Hong Kong formal prose is the hardest case. 政府發言人表示，行政長官會同行政會議今日通過有關建議 contains no character, orthographic form or lexical item that is not also ordinary Taiwanese; only 行政長官 identifies it, and Taiwanese media writes 行政長官 when reporting on Hong Kong. No word list can resolve that at sentence granularity.

Use the block gate. TWFilter::Block.judge requires a paragraph to carry a minimum density of positive Taiwan evidence, and it separates the two perfectly on the sample above. For corpus construction, judge paragraphs, not sentences.

Policy.corpus.with(foreign_topics_reject: true) rejects rather than tags non-Taiwanese institutions and toponyms. On 413 268 sentences from five native sources it removes nothing; on unfiltered .tw web text it costs roughly 1 % of genuine Taiwanese sentences, because Taiwanese journalism covers the mainland constantly. Take it when purity outranks volume.

Data statement

Following Bender & Friedman (2018). A resource that claims to identify one variety of a language owes the reader an account of which variety, decided by whom, on what evidence.

Curation rationale. Traditional-script text conflates at least four written standards. Language identifiers report zho_Hant for all of them, and script-based filtering admits Hong Kong Cantonese, Macau usage and mainland text converted from simplified. These tables separate Taiwan Mandarin from the rest on orthographic, lexical and morphological evidence, and state the evidence for every decision.

Language variety. Mandarin Chinese as written in Taiwan: cmn-Hant-TW, Glottolog mand1415. The orthographic standard is 教育部《國字標準字體》, the punctuation standard 教育部《重訂標點符號手冊》. Southern Min substrate features — the 有 + V perfective, the sentence-final particles 齁 欸 囉 — are treated as Taiwanese, not as deviations from a mainland norm.

Producer demographic. The reference corpus against which every form is graded is composed by writers and speakers in Taiwan: legislators, civil servants, journalists, lexicographers, forum users and literary authors. Age, gender and first-language distributions are not recorded and cannot be recovered from the sources. Sources with an open authorship gate are excluded: hosting in Taiwan does not establish authorship in Taiwan.

Annotator demographic. One annotator, the author, a non-native learner of Taiwan Mandarin working from published standards. This is the principal limitation of the resource. Every judgement is either taken from an official government table or checked against corpus attestation, precisely because the annotator's intuition is not a reliable substitute for either. Inter-annotator agreement has not been measured.

Speech situation and text characteristics. The reference corpus spans frozen (statutory), formal (press, reference), consultative (parliamentary, pedagogical) and casual (forum, prose) registers, written and transcribed speech, 1949 to the present with the bulk after 2000. The tables themselves are character and word inventories, not running text.

Provenance. Character charts are government open data. Lexical candidates come from OpenCC conversion tables and from hand curation; retention, banding and every attestation count are computed here. Terms in NOTICES.md.

Bender, E. M., & Friedman, B. (2018). Data Statements for Natural Language Processing: Toward Mitigating System Bias and Enabling Better Science. TACL 6, 587–604.

Licensing

Every table may be used in commercial products. Terms differ by file and are stated in NOTICES.md, which travels with this dataset:

files	source	terms
the three MOE charts	全字庫 CNS 11643, 數位發展部	OGDL v1.0 — attribution required
simplified_only.txt, the two pair tables	derived from OpenCC	Apache-2.0 — see licenses/APACHE-2.0.txt
all other tables	authored by me	MIT

Attribution strings, the Apache-2.0 §4(b) statement of modification, and the reasoning on why a character inventory is not an adaptation of a dictionary under 著作權法 art. 9(1)(3) are all in NOTICES.md. No table carries a NonCommercial, ShareAlike or copyleft term.

Attribution

Reproduce these strings verbatim somewhere a user can reach:

數位發展部，CNS11643中文標準交換碼全字庫網站
OpenCC — https://github.com/BYVoid/OpenCC, Apache License 2.0
中華民國教育部《重編國語辭典修訂本》《國語辭典簡編本》

Versioning

Releases are tagged. Pin one for reproducibility:

from datasets import load_dataset
load_dataset("taiwan-corpora/twfilter-tables", "mainland_hard", revision="v0.1.0")

MANIFEST.json carries the version and a SHA-256 per table, so a derived measurement can record exactly which tables produced it. The consuming gem exposes the same fingerprint as TWFilter.provenance.

Citation

@misc{taiwan-corpora-twfilter-tables,
  title        = {twfilter reference tables for Taiwan Mandarin variety identification},
  author       = {Patin, Den},
  year         = {2026},
  version      = {0.1.0},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/taiwan-corpora/twfilter-tables}}
}

The decision procedure that consumes these tables is archived on Zenodo as 10.5281/zenodo.21761465. Cite the tables when the data is what you used, the software when the procedure is.

Author: Den Patin, ORCID 0009-0009-6496-382X.