Translation Workflows

Translation pipelines centered on DeepL + xCOMET. Progresses from basic translation through large-scale batch processing to glossary-integrated strict translation.

Overview

The translation workflow consists of three patterns, each building upon the previous one.

Pattern	Use Case	Additional Element
Pattern 1	General technical documents	—
Pattern 2	Documents exceeding 100 pages	Batch processing
Pattern 9	Specifications/standards requiring strict terminology	Glossary management

Pattern 1: Technical Document Translation Workflow

Overview

A high-quality translation flow combining DeepL + xCOMET.

MCPs Used

• deepl-mcp - Translation execution
• xcomet-mcp-server - Quality evaluation

Flow Diagram

This diagram shows the iterative process of translation, quality evaluation, and refinement:

Skill Definition Example

The following Skill complements this pattern by defining quality thresholds and error handling procedures:

<!-- .claude/skills/translation-workflow/SKILL.md -->

# Technical Document Translation Workflow

## Quality Criteria

- Score 0.85 or above: Pass
- Score 0.70-0.85: Requires review
- Score below 0.70: Re-translate

## Error Handling

- critical: Must fix (meaning reversal, serious mistranslation)
- major: Recommended fix (unnatural expressions, terminology inconsistency)
- minor: Optional (style issues)

## Translation Settings

- formality: "more" (use formal tone for technical documents)
- Specify glossaryId if a glossary is available

Results

This workflow has proven highly effective in practice:

• Completed 180-page technical document (1.5 million characters) in one day
• Cost: approximately $12 (less than 1/100th of traditional cost)

Design Decisions and Failure Cases

Care must be taken when setting quality thresholds.

• Rationale for 0.85 threshold: Analysis of xCOMET score distributions showed that scores above 0.85 were consistently judged "acceptable quality" by human reviewers.
• Failure case: Documents heavy with abbreviations or proper nouns may receive excessive penalties from xCOMET. In such cases, combining with glossary integration (Pattern 9) is effective.

Pattern 2: Large-Scale Translation Workflow (Batch Processing)

Overview

A batch workflow for efficiently processing large volumes of translation pairs.

Flow Diagram

This diagram shows how batch processing enables efficient evaluation and targeted refinement:

Key Points

Follow these practices when implementing batch translation workflows:

• Use xcomet:xcomet_batch_evaluate for bulk evaluation
• Address only problematic sections individually
• Further acceleration possible with GPU usage

Design Decisions

• Section split granularity: Paragraph-level splitting provides the best balance. Sentence-level loses context, while chapter-level diminishes batch processing benefits.
• Low score threshold: Judge by individual section scores, not batch averages. Even if the average is high, any section below 0.70 requires attention.

Pattern 9: Glossary-Integrated Translation Workflow

Overview

A workflow that automatically extracts terminology from specifications, builds a glossary, and produces translations with consistent terminology. An evolution of Pattern 1 (Technical Document Translation), this represents a full integration pattern where a Skill orchestrates multiple MCPs.

MCPs / Skills Used

• pdf-spec-mcp - Structured extraction of specification terminology
• deepl-mcp - Glossary management and translation execution
• xcomet-mcp-server - Translation quality evaluation (optional)
• deepl-glossary-translation Skill - Defines the orchestration workflow for the above MCPs

Flow Diagram

Differences from Pattern 1

Aspect	Pattern 1 (Basic Translation)	Pattern 9 (Glossary-Integrated)
Terminology consistency	May vary between translations	Enforced via glossary
Preparation	None required	Term extraction, classification, registration
Use case	General technical documents	Specifications and standards requiring strict terminology
MCP count	2 (deepl + xcomet)	3 (pdf-spec + deepl + xcomet)
Skill	Optional (manual flow possible)	Required (orchestrating complex steps)

Concrete Example: ISO 32000-2 Glossary

Keep (acronyms): ASCII, CFF, JPEG, PDF, TLS, URI, XML ... (15 terms)
Translate:
  cross-reference table → 相互参照テーブル
  content stream → コンテンツストリーム
  null object → nullオブジェクト  ← PDF spec uses lowercase null
  indirect object → 間接オブジェクト
  ... (56 terms)

The greatest value is enforcing domain-specific terminology rules via glossary — for example, ensuring "null object" becomes "nullオブジェクト" (lowercase, matching PDF's keyword) rather than inconsistent "NULLオブジェクト" or "Nullオブジェクト".

Repository

See shuji-bonji/deepl-glossary-translation for the full implementation. Also covered in the Skill Showcase.