shuji-bonji Translation Quality Check Skill
Purpose
Ensure high-quality translations from Japanese to English for documentation and technical content. This skill provides standardized criteria and workflows for evaluating and improving translation quality.
Why This Skill?
- Consistent quality standards across all translated documents
- Automated quality gates before publication
- Clear remediation workflow for low-quality translations
MCP Tools
| MCP | Tool | Purpose |
|---|---|---|
| xcomet | xcomet_evaluate | Calculate quality score for single pair |
| xcomet | xcomet_detect_errors | Detect error spans with severity |
| xcomet | xcomet_batch_evaluate | Evaluate multiple pairs efficiently |
| deepl | translate-text | Re-translate if quality is poor |
Quality Criteria
| Score | Rating | Action |
|---|---|---|
| 0.95+ | ⭐⭐⭐⭐⭐ Excellent | Ready for publication |
| 0.85-0.94 | ⭐⭐⭐⭐ Good | Minor review recommended |
| 0.70-0.84 | ⭐⭐⭐ Acceptable | Review and improve required |
| < 0.70 | ⭐⭐ Poor | Re-translation required |
Workflow
Step 1: Identify Translation Pairs
Locate source (Japanese) and target (English) document pairs:
docs/skills/overview.ja.md ←→ docs/skills/overview.md
README.ja.md ←→ README.md
Step 2: Evaluate Quality
For single documents:
xcomet:xcomet_evaluate
- source: Japanese text
- translation: English text
- source_lang: ja
- target_lang: en
For multiple documents (5+ pairs):
xcomet:xcomet_batch_evaluate
- pairs: [{source, translation}, ...]
- source_lang: ja
- target_lang: en
Step 3: Analyze Results
If score < 0.85:
xcomet:xcomet_detect_errors
- source: Japanese text
- translation: English text
- min_severity: minor
Step 4: Remediate
| Condition | Action |
|---|---|
| Score < 0.70 | Re-translate with deepl:translate-text |
| Score 0.70-0.84 with errors | Fix specific error spans |
| Score 0.85+ with minor issues | Optional refinement |
Decision Criteria
| Condition | Action | Rationale |
|---|---|---|
| All scores ≥ 0.95 | Approve for publication | Excellent quality |
| Average ≥ 0.85, no critical errors | Approve with notes | Good enough quality |
| Any score < 0.70 | Block publication | Quality gate failure |
| Critical errors detected | Manual review required | Semantic issues |
Examples
Example 1: Single Document Check
Input:
Source (JA): MCPだけでは不十分
Translation (EN): MCP alone is not enough
Process:
- Call xcomet_evaluate with source_lang=ja, target_lang=en
- Score: 0.973 (Excellent)
- Result: Ready for publication
Example 2: Batch Check
Input:
[
{"source": "設計思想", "translation": "Design principles"},
{"source": "実践ノウハウ", "translation": "Practical knowledge"}
]
Process:
- Call xcomet_batch_evaluate
- Average: 0.95+
- Result: All pairs ready
Example 3: Low Quality Detection
Input:
Source (JA): エージェント側でのカスタマイズが困難
Translation (EN): Customization difficult agent side
Process:
- xcomet_evaluate → Score: 0.65 (Poor)
- xcomet_detect_errors → Grammar issues detected
- Re-translate with deepl
- Re-evaluate → Score: 0.92 (Good)
Anti-Patterns
❌ Skipping Batch Evaluation
Problematic:
# Evaluating 20 pairs one by one
for each pair:
xcomet_evaluate(pair) # Inefficient
Correct:
# Use batch for 5+ pairs
xcomet_batch_evaluate(all_pairs) # Efficient
❌ Ignoring Error Severity
Problematic:
Score is 0.80, no errors shown → Approve
Correct:
Score is 0.80 → Run xcomet_detect_errors
Check for critical/major errors before approval
Related MCPs
| MCP | Usage | Required |
|---|---|---|
| xcomet | Quality evaluation | Required |
| deepl | Re-translation | Recommended |