Arjuna Badger Press

People's Language — corpus-first translation

Part of the technology exposé. The engine that makes parallel editions read like people talking — and the foundation Buabantu is built on.

Parallel editions must read like people talking, not textbook flatness. People's Language (die mense se taal) is the press's name for a translation stack that routes corpus-first — human corrections outrank any model, the way eval gates outrank raw generation everywhere else in this studio.

flowchart TB
    subgraph corpus["Human corpus (SSOT)"]
        TF[translation_fixes.json<br/>Fix-a-translation programme]
        SA[sa_urban_*.json<br/>13,703 urban corpus entries]
    end
    CC[correction_corpus.py<br/>load · route · overlay_all]
    TF --> CC
    SA --> CC
    CC --> RL[real_language.py<br/>corpus-first routing + LLM fallback]
    RL --> API[api.py<br/>POST /api/real-language]
    RL --> BATCH[translate_ab.py<br/>batch regional pass]
    API --> UI[real_language.html<br/>live demo UI]
    TF --> PRESS[fix-translation.html<br/>community log]
    classDef gate fill:#1b1b1b,stroke:#d4af37,color:#fff;
    class CC,RL gate;

Routing — three steps, one invariant

Step	Route	What happens
1	`corpus`	exact match on an accepted human fix → return immediately, no LLM
2	`ai_guided`	the model is called with a BINDING CORPUS block — human entries are law, not hints
3	`corpus_overlay`	a post-AI pass replaces any wrong phrasing still in the output with the human fix

Human entries default to weight 100 — they always outrank model knowledge. The response says which path ran, so it's auditable.

The register dial

temp runs 0 (formal / scripture) → 1 (street slang). Each edition's LANGUAGES.json sets the default for its language. That dial is the seed of Buabantu's whole product thesis: the same meaning, tuned to the register a given audience actually uses.

The corpus is real

The SA urban corpus is 13,703 entries across six languages — Afrikaans, isiZulu, isiXhosa, Sesotho, Setswana, Swahili (~2,270–2,295 each). The community fix-a-translation programme feeds more in, and every accepted fix becomes binding ground truth. This is the out-data, don't out-compute strategy made literal: the moat isn't a bigger model, it's a corpus no one else has.

The faithfulness rules

Verbatim in-culture words stay verbatim. Proper names are never translated. No translator's notes the source didn't have. And — a lesson the engine learned the hard way — the length-ratio faithfulness check needs per-language-family baselines, because agglutinative Bantu languages legitimately run ~30–35% "shorter" by word count than their English source. (The story of that catch is on the Buabantu page.)

← Back to the technology exposé · productised as Buabantu.

Craft Library · Place Wiki · About the press · View this document on GitHub · Write with us

← Back to the library