taiwan-corpora/twngrams

Taiwan Mandarin web n-grams

Word 1–4-gram counts for Taiwan Mandarin (臺灣華語, cmn-Hant-TW), computed over the Taiwan slice of a large web crawl after variety filtering by twfilter 0.1.0 with the published twfilter-tables: every sentence behind these counts passed the 教育部 character-inventory gate, the simplified-character round-trip, the mainland-orthography, mainland-lexicon, written-Cantonese, Hong Kong and Singapore detectors, and block-level evidence of Taiwan-specific usage. Counts only: no document, no sentence, no context.

Intended uses

Frequency norms for Taiwan Mandarin where existing resources describe mainland usage or conflate the two varieties: lexicography and word-list construction, stimulus selection and frequency covariates in psycholinguistics, n-gram language models and smoothing baselines, segmentation and input-method vocabularies, collocation extraction, and cross-variety comparison against mainland, Hong Kong or Singapore corpora. The counts are host counts — the number of independent websites using a form — so they measure how widespread a form is across Taiwanese web authorship, not how often it is uttered.

Limitations to weigh before use: a single register (the open web as captured by Internet Archive and Common Crawl crawls), no part-of-speech annotation, tokenization fixed by one segmentation vocabulary, and the boilerplate residuals documented below. For spoken-like or edited-register norms, complement with a corpus of that register.

Provenance

Source: HPLT 3.0 cmn_Hant, quality bins 7–10 (4 116 754 documents of 113 442 082 in the language). HPLT claims no rights in the textual content and licenses its packaging under CC0; the counts here are derived, not redistributed — see NOTICES.md.

Admission gate — the union of three conditions, minus a hard exclusion:

keep  ⟸  host ∈ *.tw  ∪  html_lang ∈ {zh-TW, zh-Hant-TW}  ∪  host ∈ allowlist
drop  ⟸  host ∈ *.cn ∪ *.hk ∪ *.mo ∪ *.sg ∪ *.my            (unconditional)

2 001 161 documents admitted: 541 467 by .tw host, 1 459 694 by html_lang. The allowlist is empty — see the note at the end.

Pipeline: twpipeline stages 02–09, policy corpus, filtering by twfilter 0.1.0. Punctuation normalized to 教育部 《重訂標點符號手冊》; sentences segmented; characters checked against the three MOE standard charts and an OpenCC-derived simplified round-trip; vocabulary checked against corpus-verified mainland markers, written-Cantonese characters, Hong Kong and Singapore forms and an erhua two-window rule; blocks required to carry Taiwan-specific evidence; sentence types deduplicated; words segmented by Viterbi over a unigram cost model with EM re-estimation.

stage	in	out	lost
03_segment	2 001 161 docs	37 772 352 sentences	—
04_script	37 772 352	37 122 017	1.72 %
05_lexicon	37 122 017	36 455 533	1.80 %
06_evidence	36 455 533	29 089 952	20.20 %
07_dedup	29 089 952	12 339 761 types	57.58 %

285 516 791 tokens over 149 022 hosts.

What a count means here

Each n-gram is counted once per host, not once per occurrence. This is not a detail; it is the difference between a frequency list and a spam census.

The web corpus is dominated by templated content whose repetition is sub-sentential. The phrase 提供相關細節的諮詢服務 occurs 263 210 times inside 263 210 distinct sentences, spread over 222 900 documents and 1 948 hosts. Sentence deduplication cannot see it — every string differs. Per-document counting cannot see it — every document differs. Host capping cannot see it — the network spans 11 691 throwaway domains, none individually large. In an occurrence-counted build over the same admitted documents, 貸款 ("loan") ranked 7th among all words, which is simply false as a fact about the language.

Counting once per host asks how many independent sites use this phrase rather than how many times a template was published. For web-derived frequency this is the correct semantics regardless of spam: a generated page is not an independent act of authorship. Counted by host, 貸款 falls from rank 7 to rank 1 299, and the unigram head becomes 的 是 有 一 在 了 我 會 也 要 不 個.

Known residual. Eleven of the top hundred four-grams are fragments of the Chinese Facebook tagline 讓人們盡情分享，將這個世界變得更開闊，聯繫更緊密, carried by the social-plugin embed on roughly 3 900 independent hosts at about one occurrence each. Per-host counting cannot suppress it because those hosts are genuinely independent; it is third-party interface text and would need upstream boilerplate stripping or a stoplist. Loan-broker template phrases likewise survive where the network is wide enough — 哪裡 可以 借 錢 spans 2 475 nominally distinct hosts — so register-sensitive work should treat high-order n-grams carried by commercial boilerplate with care.

Files

file	rows	contents
token_1gram.tsv	35 149	hosts \t token
token_2gram.tsv	604 187	hosts \t token token
token_3gram.tsv	292 666	hosts \t token token token
token_4gram.tsv	101 914	hosts \t token token token token

Sorted by count descending, then by key, LC_ALL=C collation. hosts is the number of distinct hosts on which the n-gram occurs. Tokens are separated by a single space; columns by a tab. UTF-8, LF.

Script purity. Across all 1 033 916 published n-grams: zero simplified-only characters, zero characters of mainland traditional orthography, zero characters of written Cantonese, zero Hong Kong or Singapore word forms, and zero Han characters outside the three MOE standard charts, verified against twfilter-tables after counting. One bigram, 信 息 at 42 hosts, is a segmentation artefact of the documented 通信息 / 可信息 exception rather than the mainland word 信息.

Publication floor

No n-gram below 40 hosts is published. Combined with sentence-type deduplication, that means 40 independent sites, not 40 repetitions of one template — stricter than the Google Books Ngrams rule of 40 books — and it makes reconstruction of any source document impossible by construction rather than by assertion.

Known properties of the source

Gate quality. Unique Taiwan-attested sentence types produced per admitted document: .tw host 7.81, html_lang on a generic TLD 5.56. A 1.41× gradient, not a cliff. The html_lang gate supplies 65.7 % of the corpus at about two thirds of the per-document yield.

Host concentration. Top 10 hosts hold 8.8 % of sentence types, top 100 20.8 %, top 1 000 46.3 %, top 10 000 82.8 %. The single largest is chinatimes.com at 3.3 %. Per-host counting neutralizes this for the published figures.

Allowlist. 300 candidate hosts were derived automatically from the 1 927 814 documents that failed both gates, and then not used. Ranked by evidence density the list is headed by localized zh-TW product pages of international vendors, which pass every origin test because localization vendors use correct Taiwanese vocabulary, but which are translated rather than natively composed. They would have added 0.3 % more volume.

Reproduction

twpipeline 01_ingest --input cmn_Hant/7_1.jsonl.zst --residue residue.jsonl > kept.jsonl
twpipeline run --from 02_normalize --to 07_dedup --drop --out staged < kept.jsonl
twpipeline run --from 08_tokenize --to 09_count --vocabulary segvocab.json \
  --orders 1,2,3,4 --floor 40 --once-per host --out ngrams < staged/07_dedup.jsonl

Use the staged form, not a shell pipe: with six processes on a pipe the whole chain runs at about 3.5 of 18 cores because every stage blocks on a 64 KiB pipe buffer, while the staged form reaches 8–11 cores per stage.

The full build runs on a single MacBook Pro (Apple M5 Max, 12 of 18 cores used, 128 GB unified memory) in under two hours wall time, peaking at 47.9 GB resident during the external-sort count. Per-stage wall time, CPU seconds, parallelism and peak memory are recorded in twpipeline's BENCHMARKS.md.

Position among Taiwan Mandarin corpora

The corpora usually cited for Taiwan Mandarin are the web corpus TaiwanWaC and, for speech, the Sinica Taiwan Mandarin Conversational Corpus (TMC) and the NCCU Corpus of Spoken Taiwan Mandarin. The spoken corpora are complements, not alternatives: TMC records 43 hours of conversation from 170 speakers (397 693 lexical words), NCCU a smaller set of face-to-face conversations in discourse-analytic transcription, and neither is a source of web-scale written frequency. The direct comparison is TaiwanWaC, and the differences are of design rather than size:

	TaiwanWaC	twngrams
volume	260 M words	285 516 791 tokens over 12 339 761 sentence types
what a frequency is	occurrences in the crawl	independent hosts using the form
variety control	web collection of traditional-script text; no text-level filter published	per-sentence script, orthography, lexicon and evidence checks; per-stage losses published
post-hoc verification	—	zero simplified-only, mainland-orthography, written-Cantonese and HK/SG forms across all 1 033 916 published rows
deduplication	document-level (Corpus Factory pipeline)	exact sentence-type dedup (57.58 % of sentences removed) plus ≥ 5-host boilerplate flagging
access	Sketch Engine subscription, query interface	CC0 files, direct download
rebuildability	search-engine collection, not repeatable	versioned public crawl, published pipeline and per-stage benchmarks

Two of those rows are measured claims, not positioning. Variety: language identification cannot separate the four written standards that share traditional script, and URL metadata is not sufficient either — in this very corpus, sentences already admitted by a .tw-host or html_lang: zh-TW gate still lost 1.72 % on character inventory, 1.80 % on lexicon and 20.20 % on block-level Taiwan evidence. A corpus assembled from Taiwan URLs without a text-level filter retains exactly that material. Counting: occurrence counts inherit host concentration — the largest single host here holds 3.3 % of sentence types and the top 100 hold 20.8 % — and sub-sentential template networks inflate single phrases by two orders of magnitude (the 135× case documented above). Host counts are immune to both by construction.

What TaiwanWaC has that this dataset deliberately does not: full sentence context, concordances, word sketches and PoS annotation. Work that needs KWIC lines should use TaiwanWaC or rebuild from a crawl with twpipeline; work that needs open, redistributable, dispersion-based frequency norms is what these tables are for.

Citation

@misc{taiwan-corpora-twngrams,
  title        = {Taiwan Mandarin web n-grams},
  author       = {Patin, Den},
  year         = {2026},
  version      = {0.1.0},
  doi          = {10.57967/hf/9810},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/taiwan-corpora/twngrams}}
}

Origin filtering by twfilter, archived as 10.5281/zenodo.21761465; its reference tables are published as taiwan-corpora/twfilter-tables.

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