How does it work?

The ontology

At the foundation sits a set of thousands of precisely defined, semantically simple concepts — drawn from research into universal semantic primitives and a well-established core vocabulary. Every concept means exactly one thing, categorized into one of seven semantic groups, so a word that's ambiguous in English never gets a chance to smuggle that ambiguity downstream.

Try it yourself — search "be" on the live Ontology and see all the distinct concepts English flattens into one word.

Semantic representation

The source text is re-expressed as simple propositions built from those concepts, richly annotated with the linguistic features a translator would actually need — things like number (a language can distinguish singular, dual, trial, or plural in ways English never marks), who's tracking whom across a passage, and relative proximity. Where scholars read a passage differently, the representation can hold more than one reading.

Try it yourself — see Genesis 6:8's full representation on the live Sources app; hover any tagged word for its features.

The target lexicon

Every word in a target language is entered once, in one of seven syntactic classes — nouns, verbs, adjectives, adverbs, adpositions, conjunctions, particles — along with the features and forms a linguist defines for that language. From there, the rules for inflection and word formation are handled automatically.

In Tagalog, entering the verb alaga ("to care") once generates its full paradigm automatically — nag-alaga, nag-aalaga, mag-aalaga for actor-focus forms, inalagaan, inaalagaan for object-focus forms — without a linguist writing out every inflected form by hand.

Transfer grammar

This is where a semantic representation stops being language-neutral and starts becoming specific to your language: restructuring propositions, resolving how verbs take their arguments, working out relative clauses, noun relationships, and the collocations that don't translate word-for-word.

In Tagalog, when two adjacent noun phrases share the same possessor, a rule collapses the repeated possessor marker on the first noun rather than stating it twice — the way a fluent speaker actually says "the king's food and wine," not "the king's food and the king's wine."

Synthesizing grammar

The final stage generates actual text — resolving agreement, choosing the right word for the context, handling affixes, setting word order, applying sound changes, and marking pronouns — to produce a complete draft.

Word order

A phrase-structure rule fixes where each kind of constituent is allowed to sit in a clause — in Tagalog, a post-verbal adverb ordinarily precedes a pronominal subject: Nakakakita na ako ("I can see now").

Case affixes

A single rule inserts the right case marker for a noun phrase depending on its grammatical role and whether it names a common or proper noun — in Tagalog, ang for an absolutive common noun, si for an absolutive proper noun, ng for an ergative common noun, across a table of a few dozen combinations.

Tense affixes

In Kewa, verb tense is marked with a suffix that depends on the subject's person and number as well as the tense itself — first-person singular present takes -lo, past -wa, remote past -su, future -lua — and the pattern shifts again for dual and plural subjects.

Pronoun clitics

In Kewa, a reciprocal action ("to each other") is marked with the clitic -na attached after the object noun phrase, rather than a separate word: Nííáámé níáána bipa búkú kálama ("We gave books to each other").

Sound changes

In Tagalog, the case marker ng contracts to a bound suffix -ng when it follows a vowel-final verb — sinabi ("said") plus ng becomes sinabing, not two separate words.

What that process looks like day to day for a translation team — and where AI assistance fits into it — is covered in Workflows. Or see a real example of all five layers run on a single verse.