The translation pipeline

藏文文献整理与翻译流程

How a Tibetan text travels from a scanned page to something published here. Every stage is a script under ~/tibetan-pipeline/, and the runbook below is the same one we follow by hand.

Stages

1 · Find

Identify a text worth translating: check the registry for its English status first. A text with no translation, or a partial one, is a candidate. Record its catalogue id (Tohoku or BDRC work id) so every later stage can be traced back.

2 · Ingest

Acquire the source: a digital e-text where one exists (BDRC eText, rKTs, OTDO), otherwise page images from a scan. Provenance and licence are recorded at this step, not afterwards.

3 · Transcribe or OCR — and correct

Machine OCR of Tibetan is unreliable: stacks of consonants and the vowel-sign tsheg are easy to confuse, and a single wrong character changes the word. Two paths are supported:

• e-text — already transcribed, so verification only.
• OCR — run recognition, then correct. Corrections are flagged, never silently applied, so the difference between the source and our reading stays auditable.

Quality gate: a text is not eligible for translation until a human has signed off the transcript.

4 · Translate

Machine translation produces a first pass; every passage is then reviewed against the Tibetan. Terminology is pinned to a glossary so a term is rendered the same way throughout a text. Where an earlier English translation exists it is consulted for terminology, never copied.

5 · Publish

The translation is published here with its Tibetan source side by side, its provenance, and its review state.

Runbook

# 1. register a candidate (records catalogue id, licence, source)
pipeline/find_text.py --toh 21 --source 84000 --note "no modern translation"

# 2. ingest the source (e-text or page images)
pipeline/ingest.py --work toh21 --from bdrc-etext --out works/toh21/source/

# 3a. if images: OCR
pipeline/ocr.py --work toh21 --engine bdrc-ocr --out works/toh21/ocr/

# 3b. correct the transcript (human review; diffs are kept)
pipeline/transcribe.py --work toh21 --review --out works/toh21/text/

# 4. translate, with the glossary pinned
pipeline/translate.py --work toh21 --glossary glossary/ --model local
pipeline/review.py --work toh21 --against-source

# 5. publish
pipeline/publish.py --work toh21

Status of each stage