README.md

metadata

license: cc0-1.0
language:
  - hak
  - en
  - id
pretty_name: Hakka (Pontianak & Singkawang, West Kalimantan) Phrasebook & Vocabulary
size_categories:
  - 1K<n<10K
task_categories:
  - translation
  - text-generation
tags:
  - hakka
  - low-resource
  - indonesia
  - west-kalimantan
  - pontianak
  - singkawang
  - phrasebook
  - language-learning
  - parallel-corpus

Hakka (Pontianak & Singkawang, West Kalimantan) Phrasebook & Vocabulary

A structured dataset of West Kalimantan Indonesian Hakka (客家話) vocabulary, phrases, and grammar patterns, romanized and glossed in English (and Indonesian for the parallel-sentence corpus category) — including, where the source material allows it, a dialect-level distinction between Pontianak Hakka and Singkawang Hakka, two genuinely different varieties often lumped together under one generic label.

Hakka is spoken by roughly 80 million people worldwide but is under-documented online, especially the Indonesian varieties. This dataset focuses specifically on the Hakka spoken by the Chinese Indonesian (Chindo) community of West Kalimantan (Pontianak, Singkawang, Sanggau).

Content

3,204 entries across four categories:

category	rows	description
corpus	1,985	Full parallel sentences, Indonesian ↔ Pontianak Hakka, from a personal phrasebook-style corpus (travel, daily life, services) — the largest and most naturalistic part of the dataset.
vocabulaire	748	Single words: numbers, family terms, body parts, health/illness, food, verbs, adjectives, weather, places, colors, etc. Includes words extracted from the corpus sentences, a separate personal vocabulary list, and a thesis-sourced companion dictionary — see Source & compilation and Limitations.
communication	339	Everyday phrases and dialogue: greetings, shopping, work, small talk, register/politeness variation.
grammaire	132	Grammar patterns (question formation, tense/aspect markers, connectors, numeral classifiers) with example sentences.

(Row count dropped from a peak of 3,494 on 2026-09-02 after a cleanup pass: the source vault's markdown files had accumulated duplicate/near-duplicate entries across four separate append passes, and one file — a "quick reference" cheat-sheet that re-packages entries already documented elsewhere — was found to have moved into a scanned folder and started contributing ~90 duplicate rows; both are now deduplicated/excluded. No unique vocabulary was intentionally removed.)

The vocabulaire/communication/grammaire categories come from personal study notes; corpus comes from a separate, larger personal compilation (see Source & compilation).

Fields

field	type	description
id	int	Row identifier.
hakka	string	The Hakka word/phrase/sentence, romanized (no standardized romanization system exists for this dialect — see Limitations). May carry an inline dialect tag like [PTK] — see Dialect notation below.
french	string or null	Legacy field, always null as of 2026-09-03 — the source vault was fully translated from French to English on that date, see Source & compilation. Kept for schema stability.
english	string or null	English gloss, when available (vocabulaire/communication/grammaire only).
indonesian	string or null	Indonesian source sentence (corpus rows only — this is the pivot language for that part of the dataset, not a translation of english).
notes	string or null	Dialect/usage notes from the source material.
category	string	vocabulaire, communication, grammaire, or corpus.
subcategory	string	Topic within the category (e.g. nombres, salutations, structures-de-phrase, khek-pontianak-phrasebook).
section	string or null	Original subsection heading, for extra context.
dialect	string or null	Parsed from the dialect tag (see below), one of pontianak, singkawang, west-kalimantan (generic, source doesn't distinguish), indonesia-general, bangka, taiwan, china-mainland, malaysia, universal, variation-noted — or null if the source material carries no tag at all (mostly older, untagged entries; not a claim that the word is dialect-neutral). All corpus rows are pontianak.
extra	object	Any additional columns from the source table that didn't fit the fields above (e.g. Mandarin/Malay comparison terms, register form, grammar pattern description, Chinese characters for Singkawang loanwords). Empty object if none.

corpus rows use indonesian instead of english, since the source material is an Indonesian↔Hakka phrasebook rather than an English-glossed one.

Dialect notation

dialect is parsed automatically from inline bracket tags in the source notes (either on the term itself, e.g. asuk [PTK], or on the table/section heading it belongs to, e.g. ## Singkawang [SKW]). Full tag legend:

• [PTK] — Pontianak Hakka (坤甸客語) — sourced from Huang Su-chen's 2013 master's thesis and community sources explicitly claiming Pontianak.
• [SKW] — Singkawang Hakka (山口洋客語) — sourced from Huang Hui-chen's 2008 master's thesis and a 2023 peer-reviewed article (verified with two native Singkawang speakers).
• [WK] — West Kalimantan, generic/unspecified (the source doesn't distinguish Pontianak from Singkawang — most of the pre-existing vocabulary in this dataset falls here or has no tag at all).
• [ID] — General Indonesian Hakka · [BK] — Bangka dialect · [TW] — Taiwanese Hakka · [CN] — Mainland China Hakka · [MY] — Malaysian Hakka · [Universal] — used across all Hakka dialects · [Variation] — a regional variation exists without further detail.

Pontianak vs Singkawang, briefly: despite sharing the generic "West Kalimantan Hakka" label, these are two distinct varieties — different Chinese-mainland origin (Meixian vs. Hailu/Hopo/Wuhua), different local nicknames ("parler doux" 軟話 vs. "parler dur" 硬話), and different tone systems (Pontianak: 24/11/31/55/2/5 · Singkawang: 53/55/24/31/34/5), while still sharing a layer of recent loanwords (Teochew, Cantonese, Malay/Indonesian).

Caution: the real dialect continuum is fuzzier than these tags suggest. Concrete example: fung means "smoke/enfumer" in Pontianak Hakka (tone 4) but "wind" in Singkawang Hakka — and a third, unrelated community source glosses it as "red". Similarly, "to eat rice" is set fan/sit fon in Meixian-leaning Pontianak sources, but shit fon is claimed by some as the Singkawang/Bangka form — while other self-identified Pontianak sources also use sit fon. Treat [PTK]/[SKW] as "what the source declares", not as a settled norm — this romanization also does not capture tone at all.

Source & compilation

The source notes were originally written in a mix of English and French; on 2026-09-03 the French portions (headers, prose, glosses) were translated to English throughout, since the intended audience is primarily English-speaking. The french field is a leftover of that earlier state and is now always null — kept in the schema rather than removed, so existing code that reads it doesn't break.

This dataset was compiled from personal language-learning notes. Most of the vocabulaire/communication/grammaire categories draw on a mix of family/community knowledge and publicly available Hakka-learning blogs and dictionaries for the West Kalimantan dialect — that portion is not an academic corpus, treat it as a community/hobbyist phrasebook, not a normative reference. The [PTK]/[SKW]-tagged subset within it is more rigorously sourced: two Taiwanese master's theses (Huang Su-chen 2013 on Pontianak Hakka, Huang Hui-chen 2008 on Singkawang Hakka, both National Central University) and a 2023 peer-reviewed article (Semiotika, Universitas Indonesia) that verified its Hakka loanword list with two native Singkawang speakers. The theses themselves are institutionally locked in Taiwan; the original (2026-08-21) batch of this dataset relied on their Wikipedia summaries, not the primary theses.

A larger, more rigorous batch (2026-09-02, ~266 rows) was added from a companion dictionary built by a separate compilation effort that went back to the same two theses directly rather than their Wikipedia summaries, with per-entry Chinese characters, IPA tone values for part of the Singkawang entries, and an explicit 3-tier confidence system (★★★★★ = directly attested in a thesis, ★★★★☆ = attested in an explicit community dictionary/blog, ★★★☆☆ = inferred from the corpus category, unverified) — carried into this dataset's notes field per row. That compilation cites additional source codes (S3/S4/S5/S6) that are never defined in its own bibliography — an unresolved gap in the source material, not something this dataset can fix.

The corpus category (1,985 rows) comes from a separate personal phrasebook-style compilation built over several years (first assembled 2015, last revised 2022), independent from the notes above — likely originally translated from a common multilingual travel-phrasebook template (many sentences follow a "duty-free/hotel/directions/health" pattern typical of that genre), then rendered into Pontianak Hakka. It has not been cross-checked against the rest of this dataset for spelling consistency (e.g. it tends to write some compounds as one word — focha — where the vocabulaire category spells them with a space — fo cha).

A further batch (2026-09-02) was added from a third personal source: an unstructured vocabulary list and a song lyric with line-by-line Indonesian translation. This batch is untagged for dialect (no [PTK]/[SKW]) since it carries no city attribution of its own, unlike the academically-sourced subset above.

Limitations

• No standardized romanization. Spelling reflects common usage in West Kalimantan Hakka communities and varies between entries (e.g. jit/yit for "one").
• No audio or tone marking. Hakka is tonal; this dataset is text-only and does not capture tone — even the [PTK]/[SKW] tone-value note above is descriptive metadata, not annotated per-row.
• Mostly single-contributor compilation, not cross-validated by multiple native speakers for the bulk of the data — expect some inconsistencies or errors outside the academically-sourced subset.
• dialect: null is not a claim of dialect-neutrality — it just means the source material this row came from predates or doesn't use the tagging convention.
• Two romanized forms can collide with different meanings across sub-sources even within the same tag (e.g. asuk = "uncle, father's older brother" in the Pontianak list vs. the loanword a-shuk romanized asuk = "uncle, father's younger brother" in the Singkawang table) — a residual ambiguity in the underlying sources, not resolved by this compilation.
• A vocabulaire row glossing si as "lazy" (from the untagged personal-list batch) has no corroboration in the thesis-sourced companion dictionary, which documents si = "die"/"four" (context-dependent) and gives a different word entirely for "lazy" in Singkawang (lan-shi, 懶屍) — treat the "lazy" sense with real caution pending native-speaker confirmation.
• The corpus rows have not been individually spot-checked for translation accuracy, and spelling conventions differ slightly from the rest of the dataset (see Source & compilation).
• 17 vocabulaire rows (11 on 2026-08-23, 6 more on 2026-09-02, subcategories mots-essentiels, verbes, transport-vetements, lieux) were extracted from the corpus sentences by frequency analysis, with meaning guessed from context across multiple example sentences rather than confirmed by a source or native speaker — flagged as unverified in their notes field, treat with extra caution. A second frequency pass after the 2026-09-02 vocabulary batch found diminishing returns (252 unknown tokens at ≥3 occurrences vs. 280 before, and the remainder is mostly grammatical particles/classifiers too ambiguous to gloss confidently from context alone) — not exhaustively mined further.
• Unresolved conflict on akhiu: the academically-sourced Pontianak material glosses it as "uncle, father's younger brother" (paternal side); a separate personal vocabulary list glosses the same word as maternal-side uncle. Both are the same contributor's own sources and they disagree — noted in Famille.md, not resolved here.
• Several romanized forms are tone-dependent homonyms with unrelated meanings in different rows — e.g. si = "die" (verb) vs. si = "lazy" (adjective), ng = "5" vs. ng = negation particle. Both senses are kept as separate rows; this dataset cannot disambiguate them without tone marking.
• Moderate scale (~3k rows) — good enough for phrasebook/reference use and as MT seed/fine-tuning data for a low-resource pair, but still too small to train a translation system from scratch.

Loading

from datasets import load_dataset

ds = load_dataset("rwm1cr0/hakka-khek", split="train")
ds = ds.filter(lambda r: r["category"] == "vocabulaire")

License

Released under CC0 1.0 (public domain dedication) — use freely, no attribution required.