Verbarie · methodology
How a bilingual dictionary is built
This page describes the working order: where the data comes from, how it turns into a dictionary entry, what is verified before a release, and what happens to a translation nobody is sure about. The same order applies to every next language pair — it is not about Russian, it is about method.
The working order
Seven steps, the last of which never ends
The order is not invented here. It follows the lexicographic cycle as it is normally practised: gather facts about words → organise them into entries → publish → maintain.
- 1
Plan
Who the reader is, which pair, which formats, which licence. This is also where headword policy is decided: what becomes an entry at all.
- 2
Collect sources
Six data roles: headwords and parts of speech, frequency, morphology and word forms, pronunciation, bilingual translations, examples. No role may stay empty.
- 3
Compile
Headwords → senses → glosses. Forms, transcription, examples and idioms are attached to the sense, not to the word.
- 4
Edit and check
Two-sided consistency checks, sense hygiene, coverage thresholds, adjudication of the disputed cases.
- 5
Build the files
Device generation: index, collation, anchors. Separately for MOBI, Kobo, StarDict, Apple Dictionary and EPUB.
- 6
Test on the device
Tap-lookup on a real e-reader: tail forms, a word at the start of a sentence, regressions from earlier releases.
- 7
Release and maintain
A dictionary is never finished: corrections, new senses and fresher frequency data send the work back to step 3. This is where most of the work lives.
Sources
Open licences only — and why that is not charity
Nothing enters the dictionary that cannot be legally used and passed on. This is not an ideological position but a survival condition: a dictionary assembled from someone else’s proprietary data can be neither released, nor corrected, nor handed to another person.
| Data role | Source | Licence | What for |
|---|---|---|---|
| Frequency and word forms (FR) | Lexique | CC BY-SA | Decides which words enter the dictionary at all, and which forms will be recognised. |
| Extra vocabulary, IPA, synonyms | Wiktionary | CC BY-SA | Picks up what frequency lists miss: rare, regional and specialised words. |
| Literary and rare vocabulary | Littré (1863) | public domain | The classic dictionary of French. Needed for the nineteenth century, which is exactly what people read on e-readers. |
| Russian morphology | OpenCorpora | CC BY-SA | Provides declension and conjugation: without it «городами» never finds «город». |
A consequence of share-alike: the finished dictionary also goes out under CC BY-SA 4.0. A purchased file may be copied and passed on — the payment is for the assembled and verified result, not for permission to use it.
Translations
The machine writes the draft, a person signs it
First glosses are generated by a language model — one sense at a time, with the part of speech and context given, not “translate this word”. The model handles the mass of ordinary vocabulary well and predictably badly with polysemy, register and terminology.
So the machine output counts as a draft, not a translation. It goes through automated checks, and anything disputed goes to a human. A correction is applied only when two independent signals agree: a source and a reviewer, two reviewers, or a reviewer and a native speaker.
The project’s first rule: confidently wrong is worse than missing. If the sources are silent or disagree, the sense goes into a queue — not into the dictionary, and not into a guess.
Quality control
What is checked before every release
Two-sided reconciliation
If word A is translated by word B, then B must also be findable in the other direction. A mismatch is a signal that one of the two senses was invented.
Sense hygiene
Checks that the senses of one entry do not duplicate each other, do not mix parts of speech, and do not quietly replace a translation with a definition.
Coverage thresholds
Share of entries with transcription, with an example, with gender marked. The release does not build if a threshold has slipped since the last one.
On-device testing
Automated tests cannot see what the eye sees: a truncated entry, a misplaced anchor, Cyrillic sorted wrong.
The checks work as a ratchet: no number may get worse than in the previous release. Improving one class of entries by quietly breaking another is not an improvement, and the build refuses to agree that it is.
A new pair
What it takes to add a language
The French half is already built and reused whole: frequency, word forms, transcription, conjugation, set phrases. A new pair needs three things from the second language — an openly licensed word list with parts of speech, morphology so that inflected forms lead to the lemma, and a person who reads the glosses and says where they lie.
The first two I can usually find myself. The third I cannot. Which is why a volunteer who reads in their own language and can say “I am not sure here” is the scarcest resource this project has.