What is Translation Memory?
Translation Memory (TM) is a database technology used in computer-assisted translation (CAT) tools that stores source text segments alongside their translations. When a translator works on new content, the TM automatically identifies identical or similar text that was previously translated, suggesting the existing translation for reuse.
This technology forms the foundation of modern translation workflows, enabling organizations to build valuable linguistic assets over time that improve efficiency, reduce costs, and ensure consistency.
How Translation Memory Works
Segmentation
The first step in the TM workflow is segmentation — dividing source content into translatable units. Typically, segments are:
- Sentences (most common)
- Paragraphs
- Headings
- List items
- Table cells
The CAT tool uses segmentation rules based on punctuation, formatting, and language-specific patterns.
Storage Structure
Each TM entry (called a translation unit or TU) contains:
- Source segment: Original text in the source language
- Target segment: Translation in the target language
- Metadata: Creation date, translator, project, quality status, etc.
- Context: Surrounding text, document structure information
Match Types
When analyzing new content, the TM identifies different match levels:
| Match Type | Description | Typical Discount |
|---|---|---|
| 100% Match | Identical segment | 70-90% |
| Context Match (101%) | 100% + same surrounding segments | 80-95% |
| Fuzzy Match (75-99%) | Similar but not identical | 30-70% |
| No Match (0%) | No similar content found | 0% |
Types of Translation Memory
File-Based TM
Stored as local files, typically in TMX format.
Advantages:
- Easy to share and backup
- No server infrastructure needed
- Works offline
Limitations:
- No real-time collaboration
- Version control challenges
- Manual synchronization required
Server-Based TM
Centralized database accessed by multiple users.
Advantages:
- Real-time updates
- Concurrent access
- Centralized management
- Better security controls
Limitations:
- Requires server infrastructure
- Internet connection needed
- Higher complexity
Cloud TM
Hosted in cloud infrastructure.
Advantages:
- No infrastructure management
- Automatic backups
- Global accessibility
- Scalability
Limitations:
- Data residency concerns
- Ongoing subscription costs
- Dependency on provider
TM Leverage and ROI
Calculating TM Leverage
TM leverage measures the percentage of new content that matches existing translations:
Leverage = (Matched Words / Total Words) × 100
Industry Benchmarks
| Content Type | Typical Leverage | Cost Savings |
|---|---|---|
| Software updates | 60-80% | 50-70% |
| Product manuals | 40-70% | 35-60% |
| Legal contracts | 30-50% | 25-40% |
| Marketing | 10-30% | 10-25% |
| Technical docs | 50-70% | 40-60% |
Long-Term Value
TM value compounds over time:
- Year 1: Building the asset (lower leverage)
- Year 2-3: Significant leverage from previous work
- Year 4+: High leverage, major cost savings
A well-maintained TM can reduce translation costs by 40-60% for repetitive content types.
Best Practices for Translation Memory
TM Maintenance
Regular Cleanup
- Remove duplicate entries
- Delete outdated translations
- Fix inconsistencies
Quality Control
- Review before importing external TM
- Tag entries by quality level
- Implement approval workflows
Organization
- Use descriptive TM names
- Maintain separate TMs by domain/client
- Document TM history and sources
Segmentation Strategy
- Use consistent segmentation rules
- Avoid over-segmentation (loses context)
- Avoid under-segmentation (reduces matches)
- Consider abbreviation handling
TM Integration
- Connect TM with terminology databases
- Implement concordance search
- Set appropriate penalty levels for fuzzy matches
- Configure auto-propagation settings
Translation Memory vs Machine Translation
| Aspect | Translation Memory | Machine Translation |
|---|---|---|
| Source | Human translations | AI-generated |
| Quality | Verified, consistent | Variable |
| Learning | From past work | From training data |
| Best for | Repetitive content | New content |
| Cost model | Asset building | Per-word/API |
Modern Approach: Many organizations combine both:
- Check TM first for existing translations
- Use MT for no-match segments
- Human review and post-editing
- Feed approved translations back to TM
TMX: The Standard Format
TMX (Translation Memory eXchange) is the ISO standard format for exchanging translation memories between different CAT tools.
TMX Structure
<tmx version="1.4">
<header srclang="en" adminlang="en" datatype="plaintext"/>
<body>
<tu>
<tuv xml:lang="en">
<seg>Hello, world!</seg>
</tuv>
<tuv xml:lang="ru">
<seg>Привет, мир!</seg>
</tuv>
</tu>
</body>
</tmx>
TMX Benefits
- Universal compatibility
- Preserves metadata
- Supports multiple languages
- Industry standard
Advanced TM Concepts
Context Matching
Context matching considers not just the segment itself but surrounding content:
- Previous segment: What came before
- Next segment: What comes after
- Document structure: Headings, formatting
This enables more accurate matches for ambiguous segments.
TM Penalties
CAT tools apply penalties to reduce match scores in certain situations:
- Different document type
- Different client/project
- Old translation date
- Multiple TM sources
TM Alignment
Creating TM from existing translations:
- Import source document
- Import translated document
- CAT tool aligns sentences automatically
- Human review and correction
- Import to TM
FAQ
How big should a Translation Memory be?
There's no ideal size — quality matters more than quantity. A well-maintained TM of 50,000 translation units can be more valuable than a messy TM with 500,000 units. Focus on relevance and accuracy.
Can Translation Memories be shared between different CAT tools?
Yes, using the TMX (Translation Memory eXchange) format. TMX is the industry standard that enables TM exchange between virtually all CAT tools.
How do fuzzy matches work?
Fuzzy matching algorithms compare segments using techniques like Levenshtein distance, token matching, and n-gram analysis. A 75% fuzzy match means approximately 75% of the content is identical, with 25% different.
Should I use one TM or multiple TMs?
This depends on your workflow. Common approaches:
- One master TM + project TMs
- Separate TMs by domain/client
- Tiered TMs (verified vs. unverified)
How do I handle TM conflicts?
When multiple TMs contain different translations for the same source:
- Prioritize TMs by reliability
- Use context matching
- Implement approval workflows
- Consider machine learning-based selection
What's the difference between TM and termbase?
TM stores translated segments (sentences), while a termbase stores individual terms with definitions and approved translations. Both work together in CAT tools for maximum consistency.