About Weft
Weft is a library of interlinear editions. Each page sets a passage of a classical text in its original form and, under every word, two further rows: how the word most likely sounded when the text was first written, and a literal word-for-word English gloss. Under each line come one or more published translations. Notes attach to words and lines. The library runs from the Pyramid Texts of Unas (about 2350 BC) to the General Act of the Berlin Conference (1885), in twenty-seven languages.
It is dedicated to Neith, goddess of the loom, and to the author's mother. Weft is an independent project by Skip Levens; it is not affiliated with any company, product or project that also uses the name. The weft is the thread carried back and forth across the warp; the source text is the warp, and sound, gloss and sense are the weft.
Reading a page
- The original line is set in its own script. Egyptian, Sumerian and Akkadian lines are set in transliteration, with a hieroglyphic or cuneiform row generated from it; Hebrew and Persian run right to left.
- The sound row respells each word for an English reader. Capitals mark the stressed syllable, or the raised pitch in languages with a pitch accent; hyphens separate syllables. The Sound key in the masthead explains every symbol for the scheme in use.
- The gloss row gives one English cell per source word, in the source's word order. Several English words join with hyphens, and grammar rides in the English:
of-the-guestfor a genitive,she-sawfor a verb with its person. It is a crib for following the original, not a translation. - Translations sit under each line or group of lines, each attributed with its translator and year. Where a good translation is still in copyright, the page cites it as a reference and does not print it.
- Tap a word to open its detail: the dictionary headword (lemma), its grammatical form, the gloss, its sound in every scheme with the IPA, any notes on it, and where each of those came from.
- The Loom (the settings panel) turns rows on and off, switches the pronunciation scheme, chooses translations, and can veil the gloss and translations so you read the original first and tap to check yourself. Settings are kept in your browser.
- Where you are. On a long work the masthead says which stanza (or chapter, hexagram or line) is in view, out of how many; tapping it opens a grid of them to jump from. Short pages have neither.
Every page is one self-contained file. It opens from disk without a server and prints cleanly; the the Gentium Book Plus type for the text, the IPA and the glosses is served from the site itself (SIL Open Font License), so its full set of marks and letters is always present; the fonts for Hebrew, Arabic, Chinese, Japanese, Devanagari, Tamil, runes, cuneiform, hieroglyphs and Coptic load from Google Fonts when you are online, and those scripts need them to display.
Why the editions are built like software
A printed interlinear fixes one editor's judgments in one state. Weft keeps every layer of a page as a plain text file in a public repository, under version control. Four consequences follow.
- Every judgment is attributed. A correction is a small entry that names who made it, when, and why. The page shows where each value came from.
- Machine work and human work are kept apart. The pipeline generates a first layer (the text, lemma and grammar from a treebank where one exists, the sound from rules) into
gen/. Human corrections go in a separate overlay,curated/, which wins where the two differ. The generated layer can be rebuilt at any time without losing a single human decision. - Every change is reviewable and reversible. A correction arrives as a proposal, is discussed in the open, and can be undone. The history of each reading is kept.
- Rules improve the whole library at once. When a pronunciation rule is corrected, the works in that language are regenerated from it, and the change in each generated file can be inspected before it is accepted. A specialist's correction to a rule reaches every line it governs.
What the sound row claims
The first pronunciation scheme on every page is the reconstruction Weft judges best supported of how the text most likely sounded when it was first written: Dante in the Florentine of about 1307, Luther's theses in Latin as it was read in Saxony, the Rigveda with its pitch accent. Where the evidence for a text's own period is thin, the nearest well-studied stage stands in, and the page says so: Homer is read in the reconstructed classical Attic of the fifth century BC, about three centuries after the poems took shape. Later traditions (the Erasmian Greek of schools, the ecclesiastical Latin of the church, modern Icelandic) are offered second, and the reader can switch between them.
The evidence differs greatly by language. For classical Latin and Greek it is strong and well studied. Egyptian writing records no vowels at all, so any vocalized reading is a conjecture; Sumerian's sounds are known only indirectly, through Akkadian scribal usage, and its stress and vowel length are unknown. Pages whose first scheme rests on indirect evidence say so at the top and invite specialists to correct it. Languages and pronunciation sets out, language by language, what each scheme rests on and where it is weakest.
The respelling approximates the reconstruction for an English reader. It is not a claim that the original speakers sounded like English speakers.
Draft and reviewed
Much of the library is draft. Where no treebank exists (Eddic poetry, the Bayeux captions, Hammurabi), the lemma, grammar and gloss were annotated by hand against a named dictionary, and the notes were written for this edition. All of it is marked draft on the page. Only a human reviewer lifts that status, by checking the material and recording the review under their name.
Generated material is labeled by what produced it. Material produced by rule or by a treebank is not reviewed by default either; where a value is doubtful, the corrections overlay is where the better reading goes.
Texts of living communities
Most of the library is in languages with no living community of first-language speakers, or in the languages of large literate traditions. Some texts belong to communities that exist now and have their own views on how their languages and documents are published. For texts in the languages of Indigenous peoples of North America, the first such strand planned for the library, Weft follows the CARE Principles for Indigenous Data Governance (Global Indigenous Data Alliance, 2019): Collective benefit, Authority to control, Responsibility, and Ethics. The principles were written for data about Indigenous peoples; applying them to published historical texts is Weft's extension of their scope, not a use their authors named. In practice:
- Each page states whose transcription and translation it is, and whether the person stood inside or outside the community. Many historical transcriptions and translations were made by missionaries and ethnographers from outside. Where Weft transcribes a text itself, Weft is also an outsider, and the page says so.
- Each page states the community's own published position on editions of its texts, with the source, where one can be found. Where none can be found, the page says so and names what was searched. A page does not imply an endorsement that the community has not given.
- Published secular texts come first: newspapers, constitutions and laws, council records, letters, and poetry written for print. Ceremonial texts are not set out without the community's recorded consent, and the pipeline's check refuses such a work without it.
- Readers from the community come first in corrections. Their readings take precedence over Weft's, and each page invites them.
- The method of transcription is stated, with how often it errs. Weft reads a scan of Cherokee twice, once by hand and once by machine (OCR), and settles every disagreement on the scan. In a calibration on 299 signs of the 1860 Cherokee New Testament, a hand reading made without looking at any other transcription erred on 14 signs (4.7 percent) and archive.org's OCR on 21 (7.0 percent), with different kinds of error, and no error was found where those two readings agreed; 2 signs could not be settled at the scan's resolution. Tesseract, the OCR Weft runs itself, differed from that agreed text on 6 to 7 percent of signs at its usual settings, and on up to 14 percent at a poor binarization threshold. Each Cherokee page will state the rates measured on its own scan.
Sources and licenses
- Every source text, treebank and translation is public domain or openly licensed (CC BY-SA, CC BY, CC0, and the Unicode License for the Chinese reading tables). Material under non-commercial licenses is not used. Each work's manifest records the license, the address and a checksum of every source.
- Sources are not stored in the repository. Each reader fetches their own copy with
weft acquire, which shows the licenses first and verifies the checksums. - Modern scholarship is quoted only in short phrases, attributed, in the notes. Quoted words remain under their owners' copyright.
- Weft's own data is released under CC BY-SA 4.0; the code under the MIT license.
How to take part
You do not need to use git. Each kind of contribution has a short form on GitHub that asks the right questions; a maintainer turns the answer into a correction with your name on it.
- A gloss, parse, scansion or note is wrong: the Correction form.
- A pronunciation is wrong, or a scheme should be added: the Pronunciation form. A citation is required.
- You would record a passage: the Recording form. Audio is not yet played on the pages; recordings are collected now and added with credit when it is.
- A text belongs in the library: the Propose a text form.
Each page links to the correction, pronunciation and recording forms from its foot and, where the pronunciation is reconstructed, from its top. Contributors who do use git can submit the correction directly; Contributing describes both routes.
Privacy
I count visits to see which texts interest readers. On weftlibrary.org, PostHog records page views without cookies and without storing your IP address; it is set up so that I cannot identify you from it. Nothing is recorded when a page is opened from a downloaded copy. The pages are served by GitHub Pages and load their fonts from Google Fonts, and both services receive the ordinary request your browser makes, as described in their own privacy policies. To ask what I hold about you, or to have it deleted, write to skip@weftlibrary.org.
Further reading
- Contributing: how corrections, recordings and new texts are submitted and reviewed.
- Languages and pronunciation: each language's schemes, evidence and weak points.
- The lifecycle of a text: how a text moves from selection to a published page.
- The data schema: the data formats, and the conventions for gloss, sound and notes.
- The repository and its open issues.