radmada/maltese-ocr-corpus

Real Maltese OCR Paragraph Pairs

Paragraph-level (image, text) pairs from real, digitally-typeset Maltese PDFs. Built to close the synthetic-only gap in Maltese OCR training data - see the accompanying paper (LV-ROVER-MLT, DocEng 2026) for context: the paper and the corpus-building scripts (package_for_hf.py, align_pdf_paragraphs.py, under experiments/neural_resume/corpus/) are at github.com/adamd1985/doceng2026. The frozen competition submission (the Tesseract LV-ROVER-MLT ensemble this corpus is not part of) is a separate repository, linked from the paper's Availability section.

Ground truth text is never taken from the PDF's own glyph layer directly; it is cross-checked against an independently-sourced authoritative text (EUR-Lex Formex XML, or Wikipedia's plaintext- extract API) via longest-common-substring coverage before a crop is included.

Usage

from datasets import load_dataset

ds = load_dataset("radmada/maltese-ocr-corpus", split="train")
sample = ds[0]
sample["image"]  # PIL image of one paragraph crop
sample["text"]   # authoritative transcription

Each row also carries source, match_score, and a provenance record with the exact source URL, license string, page number, and bounding box of the crop.

Composition

Total pairs: 36803

Source	Pairs	License
eur-lex	94	Commission Decision 2011/833/EU (reuse permitted, attribution required, no distortion of meaning)
wikipedia-mt	36709	CC BY-SA 4.0 (attribution + share-alike required for derived text; images are renders of CC BY-SA text, same terms apply)

How it was built

1. Fetch source documents: EUR-Lex Cellar (PDF + authoritative Formex XML text) and Maltese Wikipedia articles rendered to PDF (with the plaintext-extract API as the independent text source).
2. Render pages at ~150 DPI to match the effective resolution of real-world Maltese PDF crops.
3. Segment paragraph regions from the PDF's own block structure (positions are reliable even when a PDF's ToUnicode map is broken).
4. Align each region against the authoritative text by longest-common-substring coverage; keep a pair only above the match threshold (per-pair score kept in match_score).

Diacritic and dash coverage

Character counts across the full text pool (canary characters for tokenizer/encoder correctness in Maltese OCR):

Character	Count
ascii_hyphen	324586
ħ	139918
ie	100713
għ	68126
ż	63792
ġ	59273
ċ	47315
à	6596
Ġ	6394
è	598
ù	542
Għ	3059
Ċ	2409
ò	213
Ħ	1834
Ż	1327
ì	133
em_dash	707
Ie	34
IE	6
È	11
À	9
GĦ	5

ie/Ie/IE counts are raw substring matches, same method as għ above - they are not filtered for genuine digraph usage, so some matches may be coincidental i+e sequences rather than the Maltese long-e digraph.

Grave-accented vowels (à è ì ò ù and rare uppercase À È) are almost entirely word-final (e.g. 6,468 of 6,596 à occurrences, 211 of 598 è), consistent with their role in a small fixed lexical set (Ġesù, libertà, sigurtà-type words), not a productive spelling rule. Scope note: the DocEng 2026 competition's own dev/gold character inventory (char_set.json, 117 characters) contains only lowercase à ì ò ù - no è and no uppercase grave vowel. This corpus is broader than that benchmark alphabet: it is real Wikipedia/EUR-Lex text, not limited to the competition's gold-label character set, so it naturally contains è and uppercase grave forms that the competition submission was never required to handle.

Known limitations

• Source skew: 99.7% of pairs come from Maltese Wikipedia, so the register is encyclopedic prose; legal/administrative Maltese is underrepresented (94 EUR-Lex pairs).
• All pages are digital-born renders, not scans - no scanner noise, skew, or bleed-through.
• Alignment is automatic; match_score below 1.0 means the crop and text matched by substring coverage, not exact equality. Filter on match_score if you need stricter pairs.

License

Mixed per-source, see the composition table above. Every sample's metadata.jsonl provenance.doc_meta field records its exact source URL and license string. Attribution is required for all included sources; Wikipedia-derived text additionally requires share-alike for derived text under CC BY-SA 4.0.

Citation

If you use this corpus, cite the LV-ROVER-MLT paper (DocEng 2026, Maltese paragraph OCR competition system + post-submission study).