Yiftachashkenazi/benyehuda-prose-essays-authorship

Ben-Yehuda Prose & Essays: A Public-Domain Hebrew Corpus for Authorship Attribution and NLP Research

Hebrew (עברית) · פרוזה ומסות מפרויקט בן־יהודה: קורפוס עברי בנחלת הכלל לייחוס מחברים ולמחקר בעיבוד שפה טבעית

Curated by Dr. Yiftach Ashkenazi, Technion – Israel Institute of Technology.

A Hebrew-language corpus: original Hebrew prose and essays (19th–20th century, plus earlier works) from Project Ben-Yehuda, cleaned and split into paragraphs, with every paragraph in four text versions (as published, without niqqud, with unified punctuation, without punctuation), flags instead of deletions, and author metadata.

Thanks. Every text here was digitized, proofread and released by the volunteers of Project Ben-Yehuda (פרויקט בן־יהודה). This dataset exists only because of their work. The texts come from their public-domain dump (release 2026-03). If you use this dataset, please also credit "Project Ben-Yehuda volunteers" with a link to https://benyehuda.org.

	prose	essays
works	2,746	6,543
authors	163	223 (316 in all)
words	15.4M	15.6M
rows (body / heading / footnote)	412,780 (389,458 / 11,801 / 11,521)	259,908 (227,176 / 12,452 / 20,280)

What is included

• Original works only: no work with a listed translator or a translator credited on its page.

• Public domain only. A work is included only if all four hold:

  1. it is in Project Ben-Yehuda's public-domain dump;
  2. its page on benyehuda.org shows the single label נחלת הכלל (public domain), with the matching title and author, when checked on 2026-10-06;
  3. the death year of its author (every author, for a co-authored work) is known from Wikidata and is 1955 or earlier (the Israeli term is the author's life plus 70 years);
  4. its text carries no rights notice (such as "© כל הזכויות שמורות").

  441 works that are in the dump but fail one of these tests are left out: 339 because the author's death year is unknown, 87 because the author died after 1955, 5 co-authored works, and 10 essays with a rights notice. Works the site labels by permission or orphan works are left out as well.

• Prose and essays are separate units. Essays are the site's articles and essays genre, and every essay is a unit of its own.

Configurations

config	one row per
prose_paragraphs (default), essays_paragraphs	paragraph, heading or footnote, in reading order
prose_works, essays_works	whole work: paragraphs and headings joined by newlines, footnotes left out
authors	author, with dates and identifiers

Paragraph columns (in the order the viewer shows them)

column	content
author, title	the work's author and title
paragraph_number	position in the work. Numbers are increasing, with gaps where rows were removed
role	body, heading (a title, subtitle, chapter or section line, or a number-only line), or footnote (a footnote in its own row)
text	the text as on the site, with niqqud and the original punctuation
letter_count, word_count	letters (no niqqud, spaces or punctuation) and whitespace-separated words
genre	prose or essay
work_id, url	the work's id on benyehuda.org (unique across both genres) and its link
footnote_refs	the footnote numbers cited in this paragraph. The marks are removed from the text
footnote_number, footnote_of_paragraph	for footnote rows: the note's number, and the first paragraph that cites it
is_duplicate, duplicate_of	set on extra copies of a paragraph only (see below)
duplicate_of_work, duplicate_share	set on a work that is largely contained in a larger work (see below)
is_verse, mostly_non_hebrew, mentions_author_name, co_authored	flags (see below)
text_no_niqqud	the same text without niqqud
text_unified_punctuation	without niqqud, with one form per punctuation mark: " for all double quotes and gershayim, ' for single quotes and geresh, - inside a word, – for other dashes, … for ellipses
text_no_punctuation	without niqqud and without punctuation
source_edition, site_edition_date	the edition the volunteers typed from and the date Ben-Yehuda shows for it, both as given by Ben-Yehuda. This is the edition date, not the date of first publication
site_author_id, site_copyright_label, copyright_checked_at	the author's Ben-Yehuda id, the label on the work's page (נחלת הכלל for every row), and when it was checked

The works tables start with author, title, text (the whole work), letter_count, word_count, paragraphs and headings. The other columns follow in the same order as above.

The authors table has author, birth_year, death_year, birth_place, death_place, gender, hewiki_title (Hebrew Wikipedia), wikidata_id, nli_j9u_id (National Library of Israel) and site_author_id, all from Wikidata, matched by the Ben-Yehuda author id.

Cleaning: what was removed and what is only flagged

The rule is that nothing is deleted unless it is an error, or text that the author did not write.

Removed:

• site artefacts: title/author lines, links to related articles and to recordings, image file names, footnote marks inside the text;
• text by other people: notes by Ben-Yehuda's editors ("הערת פב"י"), editors' introductions and afterwords, captions, source lines, and works written by an editor;
• double-typed paragraphs: the same paragraph again within two rows of the same work;
• leftover site markup (SPLIT_ markers, # heading marks) and illustration credits.

Headings. Title-page and structural lines are marked role = heading even where the source page marks them as ordinary paragraphs: a repeated title or subtitle, a bracketed subtitle, dateline or epigraph source, a genre label (ספור., ציור), a byline or dedication, a section number (א, ב., II), and a short line without final punctuation that stands between full paragraphs. A short line that might still be a heading is kept as body.

Kept and flagged, so that each user can decide:

flag	meaning	prose rows	essay rows
is_verse	verse lines inside the prose, often quotations	4,806	1,635
mostly_non_hebrew	fewer than half of the letters are Hebrew (5+ words)	265	1,538
mentions_author_name	the author's full name appears in the paragraph	12	256
co_authored	the work has more than one author	121	0

Duplicates: keeping one copy with one filter

Some texts appear more than once on the site, for example a story both on its own and inside a collection, or the same piece in both genres. Only the extra copies are marked. The copy that stays unmarked is the one in the largest work.

• is_duplicate / duplicate_of: an extra copy of a body paragraph of 10+ words (ignoring niqqud and punctuation). duplicate_of gives the kept copy as "work W paragraph P". There are 1,373 such rows in prose and 1,663 in essays.
• duplicate_of_work / duplicate_share: a work at least 30% of whose text (sampled 4-word shingles) appears in a larger work in either genre, with the id of the larger work and the share. This applies to 77 prose works and 211 essays.

from datasets import load_dataset

repo = "Yiftachashkenazi/benyehuda-prose-essays-authorship"
paragraphs = load_dataset(repo, "prose_paragraphs", split="train").to_pandas()

one_copy = paragraphs[~paragraphs["is_duplicate"]]                    # one copy of every paragraph
one_copy_of_each_work = paragraphs[paragraphs["duplicate_of_work"] == ""]  # drop works contained in larger ones
units = one_copy[(one_copy["role"] == "body") & (one_copy["word_count"] > 10)
                 & ~one_copy["is_verse"] & ~one_copy["mostly_non_hebrew"]]  # e.g. prose units for authorship attribution

Known limitations

• Text version: the text is from the dump release 2026-03. Ben-Yehuda has corrected some texts since then, so about 2% of the works compared differ slightly from the live site.
• Typing errors: digitization errors are kept as they are, for example 1 for ! or ן for ו. They are rare.
• Quotations: essays often quote other writers, including whole verse lines, and only part of this is flagged (is_verse). Authorship work should keep this in mind.
• Dates: there is no reliable date of first publication. site_edition_date is the date of the edition the volunteers typed from.
• Copyright: the public-domain status was checked carefully but not by a lawyer. The test follows Israeli law. Rules elsewhere (for example in the US) can differ.

License

The texts are in the public domain. Our additions are released under CC0 1.0: the paragraph structure, flags, normalized versions and links. The author metadata from Wikidata is CC0 as well. Citation is not required, but it is appreciated, and so is a credit to "Project Ben-Yehuda volunteers" (https://benyehuda.org).

Citation

@misc{ashkenazi2026benyehudaprose,
  title  = {Ben-Yehuda Prose \& Essays: A Public-Domain Hebrew Corpus for Authorship Attribution and NLP Research},
  author = {Ashkenazi, Yiftach},
  organization = {Technion -- Israel Institute of Technology},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Yiftachashkenazi/benyehuda-prose-essays-authorship}},
  note   = {Texts from Project Ben-Yehuda (https://benyehuda.org), public-domain dump release 2026-03}
}

בעברית

פרוזה ומסות מקוריות בעברית מפרויקט בן־יהודה, שכולן בנחלת הכלל.

איך נבנה הקורפוס: הוא חולק לפסקאות לפי סדר הקריאה. כל פסקה מופיעה בארבע גרסאות: כפי שפורסמה, בלי ניקוד, עם פיסוק אחיד, ובלי פיסוק. רק טעויות וטקסט שלא נכתב בידי המחבר הוסרו. כל השאר נשאר במאגר, עם עמודות סימון.

כפילויות: רק העותקים הנוספים מסומנים, כך שעותק אחד מכל טקסט מתקבל בסינון אחד.

מה עוד במאגר: טבלת מחברים עם שנות לידה ופטירה.

תודה: תודה גדולה למתנדבי פרויקט בן־יהודה, שהקלידו, הגיהו ושחררו את הטקסטים לציבור.