mahalisyarifuddin/utusan-jawi-sentences

Utusan Melayu (Jawi) — Sentence-Level Corpus

Sentence-level corpus crawled from the Wayback Machine archive of utusanmelayu.com.my (Portal Jawi Utusan Melayu / Utusan Malaysia), the Malay-language news portal published in Jawi script (Arabic script).

19,842 sentences · 2,069 articles · 8 sections · dated 2016-06-06 → 2019-09-14

---

⚠️ Important note on the source domain

utusanmelayu.com.my no longer exists in live DNS — the domain returns NXDOMAIN from Cloudflare, Google and Quad9 resolvers, and no variant (.com, .net, .org) resolves either. The live successor site is utusan.com.my.

This dataset was therefore built from the Internet Archive Wayback Machine captures of the historical utusanmelayu.com.my, which are intact: 10,389 captures across 4,894 unique URLs (2016–2025), of which 2,072 article URLs were identified and 2,069 successfully retrieved (99.86%).

---

Why this corpus is unusual (and what was done about it)

Working with Jawi-script news required solving three problems that are not handled by off-the-shelf tooling. Each was diagnosed empirically and fixed in code (sentence_split.py, build_dataset.py).

1. ؞ (U+061E) is not a sentence terminator

U+061E Arabic Triple Dot Punctuation Mark is historically used as a full stop in Jawi, so a naive splitter would treat it as one. In this corpus it is never a terminator — it is a word-internal high-hamza glyph (the pre-1980s diphthong mark), not a full stop:

word	meaning
ڤرنياڬاٴن	perniagaan (business)
كڤاتاٴن	keputusan (decision)
سلاٴين	selain

Measured evidence: 0 of 11,614 occurrences are followed by whitespace, 0 occur at the end of a paragraph, and 0 sit in a terminator-like context. Splitting on ؞ would have shredded these words. The splitter therefore treats only ., !, ? (and Arabic ؟, ۔) as terminators — and U+061E is folded onto decomposed high hamza U+0674 during normalisation (see below).

2. Legacy-font Private-Use-Area codepoints

The 2016–2019 captures were authored in a legacy Jawi font whose glyphs were pasted into the HTML as PUA codepoints. These have no standard Unicode meaning and unicodedata.normalize('NFKC', …) leaves them untouched. They were identified from context and repaired:

codepoint	count	mapped to	evidence
U+E861	2,197	⹁ U+2E41 Reversed comma	ڤوليس تنترا بومبا = "polis, tentera, bomba" (a three-item list)
U+E856	130	⁏ U+204F Reversed semicolon	...برس ( 80kg ) مانيسن ( 30kg ) ڤيترول = itemised list
U+E853	1,406	ٴ U+0674 Arabic letter high hamza	بايق = baik, تربايق = terbaik, سرتاي = sertai (pre-1980s Jawi diphthong orthography)
U+061E	11,614	ٴ U+0674 Arabic letter high hamza	same glyph as U+E853, second encoding — see below
U+067E	4,648 in sentence text	ڽ U+06BD Arabic letter noon with three dots above (nya)	باپق = banyak, ميپق = minyak, ڤپاپي = penyanyi — see below

The punctuation repairs target the reversed comma ⹁ (U+2E41) and reversed semicolon ⁏ (U+204F) rather than the Arabic punctuation codepoints. The properly-encoded stretches of the source also carry genuine ، (U+060C) commas, which pass through untouched, so the two encodings stay distinguishable in the released text: sentence text holds 2,197 ⹁ (repaired from U+E861) next to 18,839 genuine ،, and headlines hold 10 ⹁ (across 10 titles) next to 254 ، (across 239 titles). The 4 byline commas are all genuine ،. Every semicolon in the corpus is U+E856-derived — 130, repaired to ⁏; no genuine ؛ (U+061B) occurs anywhere.

Also repaired: the Allah (U+FDF2) and ﷺ (U+FDFA) ligatures via NFKC, tatweel U+0640 (kashida justification, 2,190×), and stray soft hyphens U+00AD. All residual counts in the released data are zero.

Why U+061E is folded onto high hamza

U+061E Arabic Triple Dot Punctuation Mark is historically a Jawi full stop, so the obvious reading is that it terminates sentences. In this corpus it never does:

• 0 of 11,614 occurrences end a sentence or a paragraph
• 0 appear in a terminator-like context (never space-delimited, never adjacent to other punctuation)
• 99.8% are preceded by alef (10,710) or waw (886) — the hamza carriers
• 258 distinct words are attested with both codepoints for the same word, e.g. ياٴيت/يا؞يت yaitu, كراجاٴن/كراجا؞ن kerajaan, داٴيره/دا؞يره daerah, باٴيق/با؞يق baik

It is therefore the same triple-dot-above glyph as the legacy font's U+E853, inconsistently encoded. Both are folded onto U+0674 Arabic letter high hamza.

High hamza is always decomposed

The fold target is U+0674 ٴ (Arabic letter high hamza), not combining U+0654 (hamza above) and not a precomposed letter.

• U+0654 is the wrong character. It is a combining mark. NFKC canonically composes alef + U+0654 into أ, and a combining mark cannot sit before its carrier without attaching to the previous letter. An earlier build stored the glyph as U+0654 and shielded it across NFKC to stop that composition. The examples in this card were already written with U+0674; the data now matches.
• U+0674 does not compose. It is a spacing letter (category Lo). NFKC of alef + ٴ stays alef + ٴ, so normalize_jawi() is idempotent without a shield, including parse_article's second pass over the headline. Released sentence text has 13,020 high hamzas, 367 of them in headlines. 12,010 are preceded by alef and 985 by waw; 95 are followed by alef — the before-carrier order a combining mark would have lost.
• Precomposed high hamza is expanded. U+0675–U+0678 (ٵ ٶ ٷ ٸ) have only a compatibility decomposition, so NFC leaves them composed and NFKC expands them to carrier + U+0674. The normaliser applies that decomposition explicitly either way. Unicode's preferred spelling is high hamza then carrier; "decomposed" here means the decomposition mapping (carrier, then U+0674).

Native precomposed hamza-above letters that are actually in the source HTML (سأورڠ, كأدا, مسئله) are not high hamza and are left as أ / ؤ / ئ (3,594 / 381 in sentence text, plus hamza-below إ × 946). None of those come from the legacy triple-dot glyph.

Peh is nya

U+067E پ Arabic letter peh is not a Jawi letter — Jawi /p/ is U+06A4 ڤ. In this font, peh is the codepoint of nya. Initial and medial nya (U+06BD ڽ) is drawn with three dots below, the same shape as peh, and the legacy encoding put that glyph on U+067E.

Every peh in the archived prose is a nya misspelling, checked against the Rumi reading of the word:

archived	normalised	reading
باپق	باڽق	banyak
ميپق	ميڽق	minyak
سپوم	سڽوم	senyum (Gunung Senyum)
ڤپاپي	ڤڽاڽي	penyanyi
سباپق	سباڽق	sebanyak

4,648 sentence-text occurrences (524 word types) were rewritten. Genuine U+06BD already in the source (7,599 in sentence text) is unchanged, so the released text has 12,247 nya. Residual peh, including the presentation forms U+FB56–U+FB59, is zero. article_id / url keep the original CMS slug, which still contains peh; only normalised prose (text, title, author) is rewritten. Both substitutions are 1-for-1, so n_chars and n_words are unchanged and the article-level split is untouched.

3. Quote direction is inconsistent

Some pages use ” to open and “ to close; others do the opposite. A depth counter decides open/close by position rather than by codepoint.

4. Kaf and gaf are unified onto the keheh family

Standard Malaysian Jawi writes /k/ with keheh ک (U+06A9) and /g/ with gaf — keheh with a dot above — ݢ (U+0762). The archived pages are inconsistent: the same words appear with the plain Arabic kaf ك (U+0643) and kaf with dot above ڬ (U+06AC) as well as the keheh forms (e.g. اكن / اکن akan, ڬورو / ݢورو guru). The kaf family is folded onto the keheh family so a word has one spelling regardless of which glyph the source happened to use:

archived	normalised	mapped to	reading
ك U+0643 Arabic letter kaf	ک U+06A9 Arabic letter keheh	62,993 in sentence text	اكن → اکن akan
ڬ U+06AC Arabic letter kaf with dot above	ݢ U+0762 Arabic letter keheh with dot above (gaf)	25,628 in sentence text	ڬورو → ݢورو guru, نڬري → نݢري negeri

After the fold, sentence text holds 84,995 keheh ک (22,002 already in the source) and 25,628 gaf ݢ; residual ك / ڬ is zero. Both substitutions are 1-for-1, so n_chars / n_words and the article-level split are unchanged. As with peh→nya, only normalised prose (text, title, author) is rewritten; identifiers (article_id, url, slug, section_jawi) keep the original CMS/URL spelling — so the Galeri Foto section slug stays ڬالري-فوتو with its U+06AC.

Yeh and alef maksura are not unified. Arabic-script normalisers often fold alef maksura ى (U+0649) onto yeh ي (U+064A), but in Jawi the two are distinct letters, so both are left exactly as they are (179,618 yeh and 344 alef maksura in sentence text).

---

Sentence segmentation

part = "title" — one row per sentence of the headline. part = "body" — one row per sentence of the lead + body paragraphs, in document order.

Splitting is quote-aware, protects decimals (RM10.55), image ids (1.514882), times (4:31 PM) and thousands separators (RM2,999), and:

• re-attaches closing quotes that the archived markup strands on their own line (the HTML frequently breaks a quotation across paragraph boundaries);
• re-attaches bare list enumerators (1., 2., …) to the item they label, since the markup puts the marker on its own line;
• merges punctuation-only fragments into the sentence they belong to.

Headlines legitimately carry no full stop, so text ending without a terminator is expected and correct.

---

Split methodology — stratified 8:1:1 at the article level

• Unit of sampling = the article. Every sentence of an article lands in exactly one split, so no sentence-level leakage between train/validation/ test. (Verified: 0 of 2,069 articles span more than one split.)
• Stratified on section so each split preserves the section distribution.
• Random seed 0 — a single random.Random(0) instance, iterating sections in sorted order over article-ids sorted deterministically, so the split is bit-reproducible.
• Allocation is round-half-up per stratum: train = round(n·0.8), validation = round(n·0.1), test = n − train − validation.

Achieved:

split	sentences	share	articles
train	15,881	80.04%	1,655
validation	1,989	10.02%	208
test	1,972	9.94%	206

Per-section (articles → train/val/test): Nasional 1441→1153/144/144 · Hiburan 135→108/14/13 · Pendidikan 129→103/13/13 · Rencana 119→95/12/12 · Sukan 119→95/12/12 · Sejagat 100→80/10/10 · Bisnes 20→16/2/2 · Galeri Foto 5→4/1/0 · info 1→1/0/0.

Schema

field	type	description
sentence_id	string	<article_id>#t<n> for titles, #b<n> for body
text	string	the sentence, normalised
part	string	title | body
para_kind	string	lead | body (null for titles)
split	string	train | validation | test (dropped from the saved splits — it is the split key)
article_id	string	Polopoly article id, e.g. 1.514881
section	string	Latin section label
section_jawi	string	original Jawi section slug from the URL
title	string	article headline
author	string	byline, oleh/اوليه prefix stripped
date	string	ISO yyyy-mm-dd when parseable
date_raw	string	original byline date string
wayback_timestamp	string	Wayback capture timestamp used
url	string	original (now dead) URL
wayback_url	string	resolvable Wayback URL
image	string	lead image URL, if any
topics	list[string]	Latin subject tags scraped from the page
genres	list[string]	Latin genre tags scraped from the page
sent_index	int64	1-based index within title or body
lang	string	ms-Jawi
n_chars	int64	character count
n_words	int64	whitespace-token count

Sections

section	Jawi	articles
Nasional	ناسيونل	1,441
Hiburan	هيبورن	135
Pendidikan	ڤنديديقن	129
Rencana	رنچان	119
Sukan	سوکن	119
Sejagat	سجاڬت	100
Bisnes	بيسنس	20
Galeri Foto	ڬالري-فوتو	5
info	info	1

سجاڬت transliterates to Sejagat (world / international), confirmed by content (Nigeria, Nepal, Bandung).

Usage

from datasets import load_dataset

ds = load_dataset("mahalisyarifuddin/utusan-jawi-sentences")
print(ds)
print(ds["train"][0]["text"])

# filter by provenance
ds["train"].filter(lambda x: x["part"] == "title")

Or from the local build:

from datasets import load_from_disk
dsd = load_from_disk("dataset/hf/utusan_jawi_sentences")

# from parquet
from datasets import load_dataset
ds = load_dataset("parquet", data_files={
    "train": "dataset/parquet/train.parquet",
    "validation": "dataset/parquet/validation.parquet",
    "test": "dataset/parquet/test.parquet",
})

To reproduce the split exactly:

import random
rng = random.Random(0)
for section in sorted(sections):            # sections sorted by Latin label
    ids = sorted(articles_of(section))      # deterministic pre-shuffle order
    rng.shuffle(ids)
    n_train = max(1, int(n * 0.8 + 0.5))
    n_val   = min(int(n * 0.1 + 0.5), n - n_train)
    # ids[:n_train] -> train, next n_val -> validation, rest -> test

Coverage & quality

• 2,069 / 2,072 article URLs retrieved (99.86%); 1 permanent 404 in the archive, 2 /cmlink/* promotional containers excluded as non-articles.
• 2,064 / 2,069 articles have a parseable date; 1,830 have a byline (540 distinct authors); 1,953 carry Latin topic tags; 1,762 have a lead image.
• Median sentence: 113 characters / 19 words (p10 = 30, p90 = 195).
• Short sentences are dominated by legitimate Malay attributions (كاتاڽ. katanya, اوجرڽ. ujarnya, جلسڽ. jelasnya).

Limitations

• Temporal skew. The archive is heavily weighted to 2018–2019 (5,078 and 3,305 captures) with thin coverage elsewhere; Nasional is 70% of the corpus.
• Archive capture quality. Text is whatever the Wayback Machine stored; a handful of captures are truncated mid-sentence.
• Legacy orthography. Pre-1980s diphthong marks are stored as decomposed high hamza ٴ (U+0674) — 13,020 of them, 367 of them in headlines — which differs from modern Jawi (ء / اءي). Native أ / ئ in the source HTML are not rewritten. Legacy peh (U+067E) has been folded onto nya (U+06BD), and the kaf family is unified onto the keheh family (kaf → keheh, kaf-with-dot → gaf); yeh and alef maksura are kept distinct.
• Mixed script. Latin loanwords, acronyms and place names are embedded in the Jawi text, so this is not a pure single-script corpus.
• No sentence-boundary gold standard. Segmentation is rule-based and hand-validated on samples, not annotated.

Copyright

The articles are the copyrighted editorial work of Utusan Malaysia, owned and operated by Media Mulia Sdn Bhd (201801030285 / 1292311-H), Level 4 (East) & Level 5, Quattro West, No. 4, Lorong Persiaran Barat, 46200 Petaling Jaya, Selangor. The current terms are Terma dan Syarat.

Clause 3 of those terms limits use of Utusan Malaysia content to personal and non-commercial use. It forbids building a business or other enterprise on the Services, whether for profit or not. As between the reader and Utusan Malaysia, Utusan Malaysia is the sole owner of the text, graphics, and other content. Access does not transfer ownership. All rights are reserved. The terms are governed by Malaysian law, and disputes are subject to the courts of Malaysia.

This dataset is a research derivative of Internet Archive captures of the former utusanmelayu.com.my. It does not grant any licence beyond those terms. It is not open data, and it is not cleared for commercial use, redistribution, or training a product. Hak cipta terpelihara © Media Mulia Sdn Bhd.

Provenance

step	detail
discovery	Wayback CDX API, url=utusanmelayu.com.my&matchType=domain, 3 index pages
captures	10,389 captures → 4,894 unique URLs → 2,072 article URLs (Polopoly -1.<id> pattern)
retrieval	https://web.archive.org/web/<ts>id_/<url> (raw HTML, no Wayback chrome), 6 keep-alive workers, exponential backoff, checkpointed
parsing	BeautifulSoup + lxml; h1.article-title_title, div.articleLead, div.articleBody, span.author, span.date, ul.tag-list
segmentation	sentence_split.py (quote-aware, Jawi-aware)
split	build_dataset.py::stratified_split (article-level, seed 0)
build	build_hf.py → datasets.DatasetDict + parquet