File size: 3,599 Bytes
d4558b2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 | # Translation review rubric
Review the canonical claims, source evidence, and target-language surface
together. A reviewer must not grade fluency while ignoring whether the caption
is true of the evidence.
## Severity
| Severity | Meaning | Examples |
|---|---|---|
| Critical | Changes the training target or contradicts evidence | wrong action, object, actor, hand, negation, quantity, direction, state, outcome, or timestamp |
| Major | Drops or invents a material distinction | missing tool, source/destination, bimanual coordination, uncertainty, goal, error or recovery |
| Minor | Meaning remains usable but expression is degraded | awkward word order, punctuation, nonpreferred terminology without ambiguity |
Any critical error rejects the row. A major error requires correction and
re-review. Minor errors are counted by language and construction so systematic
problems cannot hide inside an average.
## Review dimensions
1. **Evidence fidelity:** every expressed claim is supported by its cited frame,
interval, sensor, source label, or human judgment.
2. **Action structure:** predicate, actor, patient, instrument, source,
destination, manner, goal, and result remain distinct.
3. **Hands and bodies:** ego versus other person, left/right, visibility,
contact, functional role, and bimanual coordination are preserved or marked
unknown.
4. **Objects and state:** object identity, attributes, containment, contact,
before/after state, and outcome are not merged or hallucinated.
5. **Time and granularity:** point, interval, atomic action, task phase, and
whole activity are not treated as synonyms.
6. **Communicative meaning:** observation, directive, question, uncertainty,
register, polarity, aspect, modality, and reported speech remain explicit.
7. **Language quality:** terminology, grammar, naturalness, locale, and audience
are appropriate without adding unsupported facts.
8. **Provenance:** source record, revision, evidence selectors, producer,
method, review state, and loss ledger remain attached.
## Language-specific checks
### Simplified Chinese (`zh-Hans`)
- Do not infer aspect particles, classifiers, completedness, or omitted
arguments from English.
- Check disposal/passive constructions such as 把 and 被 against semantic roles.
- Confirm that serial actions remain ordered and separate.
- Do not depend on whitespace tokenization.
### Tagalog (`tl`)
- Review verbal voice and pivot together with `ang`, `ng`, and `sa` phrases.
- Do not force English subject–verb–object order onto the target.
- Preserve actor versus patient prominence and source/destination roles.
- Tag code-switched spans instead of silently normalizing them.
### Spanish (`es`)
- Separate grammatical gender from physical-world attributes.
- Check number, determiner, person, register, and imperative morphology.
- Preserve reflexive constructions, clitic roles, and source/destination.
### Hindi (`hi`, experimental)
- Check case/postpositions, agreement, aspect, word order, and politeness.
- Do not infer gender or honorific commitments without evidence.
### English (`en`)
- Distinguish observation from imperative fragments.
- Do not use articles or pronouns to invent definiteness, identity, or gender.
## Reviewer independence
Model output is a candidate. The producing model cannot approve its own row.
High-risk or benchmark rows require two independent bilingual judgments and
adjudication of disagreements. Review records bind the source revision,
canonical claim set, exact target text, rubric version, and reviewer decision.
|