Quality Hardening Features¶

Overview¶

The SRT Translator includes advanced quality hardening features that automatically improve translation results by filtering problematic content and enforcing consistency rules. These features work transparently in the background to ensure high-quality, professional translations.

Key Features¶

1. Automatic DNT Term Filtering¶

Problem: Numeric and number-like terms in DNT lists can prevent proper localization.

Example: - Before: "300 milliseconds" in DNT → stays as "300 milliseconds" in all languages - After: "300 milliseconds" filtered out → becomes "300毫秒" in Chinese, "300 milisegundos" in Spanish

What Gets Filtered: - Pure numbers: "300", "6.7", "2025" - Number-like patterns: "300 milliseconds", "6.7 seconds", "2025 Q1" - Time formats: "1:30", "2:45" - Percentages: "15%", "25.5%" - Currency: "$50", "€100"

Benefits: - Better localization of numeric content - Improved readability in target languages - Maintains important non-numeric DNT terms (brands, names, acronyms)

2. DNT/Termbase Precedence¶

Problem: Conflicts between DNT terms and termbase entries can cause inconsistent behavior.

Solution: The system uses intelligent precedence based on term type:

Hard-preserve terms (acronyms, tech codes like "API", "GPU", "NASA") always remain untranslated
Soft-preserve terms can be overridden by termbase entries when a translation is specified

Example: - Hard-preserve DNT term: "API" → always stays "API" (termbase cannot override) - Soft-preserve DNT term: "machine learning" with termbase entry "机器学习" → uses termbase translation - DNT-only term: "S-Team" with no termbase entry → stays untranslated

Benefits: - Consistent behavior across all languages - Flexibility to translate domain terms while protecting technical codes - Clear hierarchy: Hard-preserve DNT > Termbase > Soft-preserve DNT > Default translation

3. Relevant Termbase Injection¶

Problem: Injecting all termbase entries can cause AI hallucinations and reduce relevance.

Solution: Only inject termbase entries that are actually present in the current subtitle text.

Example: - Full termbase: 50 terms across all business domains - Current subtitle: Only contains 15 relevant terms - Result: Only 15 relevant terms are injected, improving translation quality

Benefits: - Reduced AI hallucinations - More focused and relevant translations - Better performance and cost efficiency

4. Termbase Size Capping¶

Problem: Very large termbases can overwhelm the AI and reduce translation quality.

Solution: Automatic capping at 30 terms per language, prioritizing important terms.

Benefits: - Consistent translation quality - Predictable API costs - Focus on most important terminology

Technical Implementation¶

DNT Term Processing¶

def filter_dnt_terms(self, dnt_terms: List[str]) -> List[str]:
    """Filter DNT terms to exclude numeric and number-like items"""
    filtered_terms = []
    for term in dnt_terms:
        if not self._is_pure_number(term) and not self._is_number_like(term):
            filtered_terms.append(term)
    return filtered_terms

Tolerant Matching¶

def _compile_tolerant_patterns(self):
    """Compile regex patterns for tolerant matching of Latin keys"""
    for term in self.dnt_terms:
        if self._is_latin_text(term):
            # Handle space/hyphen variations and possessives
            base_term = re.sub(r'[\s\-]+', r'[\s\-]+', re.escape(term))
            possessive_pattern = f"{base_term}['s]?"
            self.tolerant_patterns[term] = re.compile(possessive_pattern, re.IGNORECASE)

Output Transparency¶

Enhanced Output Files¶

The quality hardening features provide complete transparency through enhanced output files:

dnt_terms.json:

{
  "description": "DNT terms processing summary",
  "user_provided": {
    "description": "Original DNT terms as provided by user",
    "terms": ["S-Team", "300 milliseconds", "2025", "API"],
    "count": 4
  },
  "filtered_for_translation": {
    "description": "DNT terms actually used during translation",
    "terms": ["S-Team", "API"],
    "count": 2,
    "filtered_out": ["300 milliseconds (filtered: numeric/number-like)", "2025 (filtered: numeric/number-like)"],
    "filtering_reason": "Removed numeric and number-like terms for better localization"
  }
}

termbase.json:

{
  "description": "Termbase processing summary",
  "user_provided": {
    "description": "Original termbase as provided by user",
    "languages": { "es": {...}, "zh": {...} }
  },
  "filtered_for_translation": {
    "description": "Termbase actually used during translation",
    "languages": { "es": {...}, "zh": {...} },
    "collisions_removed": {
      "es": {
        "filtered_out": ["S-Team"],
        "reason": "DNT collision"
      }
    },
    "filtering_reason": "Removed termbase entries that conflict with DNT terms"
  }
}

manifest.json:

{
  "processing_summary": {
    "dnt_terms": {
      "provided": 4,
      "used": 2,
      "filtered": 2
    },
    "termbase": {
      "provided_entries": 25,
      "used_entries": 22,
      "collisions_resolved": 3
    },
    "quality_improvements": [
      "Numeric DNT terms automatically filtered",
      "DNT precedence enforced over termbase",
      "Relevant-only termbase injection"
    ]
  }
}

Configuration¶

Automatic vs. Manual¶

Automatic (Recommended): - Quality hardening features are enabled by default - No configuration required - Best results for most users

Manual Control (Future Enhancement): - Option to disable specific filtering rules - Custom filtering thresholds - Manual override for specific terms

Best Practices¶

DNT Term Management¶

Focus on Brand/Name Terms: Use DNT for company names, product names, technical acronyms
Avoid Numeric Content: Let numeric terms be translated for better localization
Regular Review: Periodically review DNT terms to ensure they're still needed

Termbase Management¶

Quality Over Quantity: Focus on important, frequently-used terms
Language-Specific: Consider cultural and linguistic differences
Regular Updates: Update termbase as your content evolves

Monitoring Quality¶

Review Enhanced Output: Check the processing summaries after each translation
Monitor Filtering Results: Ensure the filtering is working as expected
Track Improvements: Use the transparency features to validate quality gains

Troubleshooting¶

Common Issues¶

DNT Terms Not Being Preserved: - Check if the term is in your DNT list - Verify the term isn't being filtered as numeric - Check the enhanced output files for filtering details

Termbase Entries Not Being Used: - Verify the term appears in your source text - Check for DNT collisions in the enhanced output - Ensure the termbase is properly configured for the target language

Unexpected Filtering: - Review the enhanced output files for filtering details - Check the processing summary for counts and reasons - Verify the filtering rules are working as intended

Getting Help¶

Check Enhanced Output Files: The detailed JSON files contain all the information about what happened
Review Processing Summary: The manifest.json shows high-level filtering results
Check Logs: Detailed logs show the step-by-step processing
Contact Support: If issues persist, the enhanced output files provide all the context needed for troubleshooting

Future Enhancements¶

Planned Features¶

Custom Filtering Rules: User-defined filtering criteria
Filtering Analytics: Historical tracking of filtering effectiveness
Quality Metrics: Automated quality scoring based on filtering results
Advanced Matching: More sophisticated pattern matching for complex terms

Contributing¶

The quality hardening features are designed to be extensible. If you have ideas for additional filtering rules or quality improvements, please contribute to the project!

Conclusion¶

The quality hardening features in the SRT Translator provide automatic, transparent improvements to translation quality while maintaining complete visibility into what's happening. These features work together to ensure professional, consistent translations that respect your brand and terminology while providing the best possible localization for your audience.