Urdatorn/linder-walberg

Linder & Walberg, Svenskt-grekiskt lexikon (1862)

Page images and a proofread transcription of Carl Linder and Carl Walberg's Swedish–Ancient Greek dictionary (Uppsala 1862, 535 printed pages + front matter): Swedish headwords in 1862 spelling, polytonic Greek equivalents in an italic Greek face, roman and spaced type, sense numbering, dense abbreviation. A hard, real-world test of multilingual, mixed-script OCR of critical and lexicographic printing, with two layers of text for every page:

• diplomatic: what the page prints, letter for letter, accents and breathings included, the printer's misprints kept;
• corrected: the same text with the editorial corrections applied (the book's own Tryckfel list, misprints found in proofreading, and orthographic fixes of the Greek), aligned strip by strip with the diplomatic text.

Configurations

config	rows	one row per
pages (default)	540	scanned page: image, raw OCR, diplomatic and corrected text
strips	3,267	strip (a column section or heading) with its box and crop
entries	26,214	dictionary entry (pages 5–539), diplomatic and corrected
corrections	9,791	finding from proofreading and the formal checks, with its decision

pages: page (scan number, 0001–0540), printed_page (scan − 4 for the dictionary pages), section (front_matter, dictionary, back_matter), image (600 dpi bilevel PNG, from the Group 4 TIFF), text_diplomatic, text_corrected, text_ocr_raw (the vision model's output before any cleaning, in its pseudo-Markdown), proofreading_corrections (how many changes the proofreader made on the page).

strips: strip (NNNN_SS), kind (col, span, page), column (1, 2, or 0 for full width), box (left, top, right, bottom in scan pixels), image (the crop), continues (the strip opens mid-entry), text_diplomatic, text_corrected. This is the unit the text was proofread in, so image and text match exactly; the strips are the natural training pairs.

entries: headword, page, strips (where the entry was read), text_diplomatic, text_corrected.

corrections: category (errata: the book's Tryckfel list; auto: mechanical fixes; misprint: reported by the proofreaders; diacritic, rules, letter: formal checks of the Greek; headword, uncertain, unknown: recorded, not changed), printed, proposal, decision, status (applied, kept, overlap, not_found), page, strip, entry, note, context.

Text conventions

• One paragraph per dictionary entry, the whole entry on one line; entries separated by a blank line. In strips, continues marks a strip whose first paragraph continues the entry above it.
• Words hyphenated at line ends are joined; real hyphens (prefixes and endings cited as such, ἀπο-, ἐκ-; Swedish suspended compounds) are kept.
• Letterspaced type is written without the spaces; no markup for italics, bold or spacing. Unicode NFC.
• Running heads and page numbers are not transcribed.

How it was made

1. OCR: each page was cut into strips at blank rows (ocr/segment.py) and read by CLLG/Qwen3VL-8B-synth_real, a Qwen3-VL model fine-tuned for Greek and Latin critical editions, with a prompt naming the book's Swedish and a logits constraint that forbids Greek letters in headwords (the model otherwise read the bold Swedish headwords as Greek, "Ηάρφιν" for Hårfin). That is text_ocr_raw.
2. Cleaning: markup removed, and a word-by-word alignment with Project Runeberg's 2013 OCR of the same scans, which reads Swedish and punctuation well and Greek badly, restored Swedish words read as Greek, the 1862 spellings the model modernised, and the book's .: separator.
3. Proofreading: every strip of every page was compared with the scan at 450 dpi by an AI agent (Claude), word by word and accent by accent, with hints from a lexicon check of the Greek. On the five pages that have a human-verified reference (0005–0009) the character error rate fell from 1.86% (raw OCR) to 1.03% (cleaned) to 0.14% (proofread); doubtful readings on those pages were settled against a physical copy.
4. Corrections: formal checks of the Greek (breathings, accents, the grave rule, accent position, attestation in a 6.5-million-form Ancient Greek corpus and in LSJ) and the proofreaders' misprint reports gave the editorial layer; the corrections config records every finding.

The diplomatic text has been read once against the scans and is not free of errors: residual rates are of the order of the 0.14% measured on the reference pages, mostly accents and breathings. The uncertain rows in corrections list readings the proofreaders could not settle from the scan.

Source and licence

The book is in the public domain. The scans are Project Runeberg's (runeberg.org/svegrek), released there without restriction. The transcription, corrections and this packaging are dedicated to the public domain (CC0 1.0).

Prepared by Albin Thörn Cleland (Lund University) for a macOS Dictionary.app edition of the lexicon; code at github.com/Urdatorn/linder-walberg.