Translation method

How we translate web novels

Translation principles

1. Translate from the original

Every translation on Ranobeki is produced directly from the original-language text. For Chinese novels, we source raws from publicly available mirrors. For Korean and English originals, we work from the author's published text.

Third-party translations are used only as private verification references — they are never published on the Platform.

2. Respect the source

Our translation pipeline is built around several core rules:

• Proper nouns preserved: Names, places, and titles are kept consistent with the community-accepted translations for each language pair.
• Cultural context kept: Concepts rooted in the source culture are explained, not erased.
• Tone-matched: The register and style of the original are carried into each target language.
• Glossary-driven: A per-series glossary anchors terminology so character names, techniques, and world-specific terms stay consistent across hundreds of chapters.
• Human-readable first: The primary goal is a natural, readable translation. Literal accuracy serves readability, not the other way around.

Every translated chapter passes through a critic stage that flags style breaks, terminology drift, and mechanical errors before publication.

3. Multi-language in one pass

Instead of translating separately into each target language, our pipeline sends the source chapter once and produces all six platform languages in a single LLM call:

• Token-efficient: The source is processed once, reducing cost and latency.
• Contextually consistent: All language outputs share the same source understanding, reducing cross-language drift.
• Six languages: Russian, English, Spanish, Portuguese, Indonesian, and Vietnamese.

Translation process

Bootstrap phase (per series)

Before translating a single chapter, the pipeline scrapes the wiki, builds a glossary of proper nouns and world-specific terms, and verifies the glossary against community reference translations. This ensures every chapter starts with a solid terminology foundation.

Bulk translation loop

Once the glossary is built and verified, the pipeline enters a streaming loop:

• Fetch: Retrieve the next chapter in the original language from the source mirror.
• Translate: Send the chapter text with the glossary to the LLM; receive all six target languages.
• Critic: A separate model reviews each language output against the original and the glossary, scoring quality.
• Fix: Languages scoring below threshold are re-translated individually.
• Export: Approved translations are exported to the frontend fixtures directory.

Escalation path

Chapters that fail critic review after two re-translation attempts are flagged for human review. Community contributors can suggest corrections through the platform interface.

AI models

The translation pipeline currently uses these models:

• DeepSeek-V4: Primary translation model — produces all six target languages from the source text in a single call.
• GLM-4.7: Critic model — reviews each language output for terminology consistency, style, and mechanical errors.
• Open-source fallbacks: Smaller open-weight models serve as fallbacks when primary models are unavailable.

Feedback and corrections

Translation quality improves with feedback. If you spot a terminology inconsistency, a mistranslated name, or a passage that reads unnaturally, you can suggest a correction directly from the reader toolbar.

Corrections are reviewed, applied to the glossary where appropriate, and back-propagated to affected chapters on the next pipeline pass.

Good translation is invisible. Great translation feels like it was written in your language.