Hybrid Cascade Architecture: Engineering Practice of Term-Aware Machine Translation

Core Information

• Project Name: Hybrid Cascade Architecture-based Term-Aware Machine Translation System
• Core Components: MarianMT Local Inference + Translation Memory Caching + Gemini 2.5 Post-editing
• Key Results: Term accuracy increased from 36.67% to 72.88% (no model retraining)
• Source: GitHub project (maintained by FatmaElMahdi1000, released on June 3, 2026)
• Reference Paper: Domain Terminology Integration into Machine Translation: Leveraging Large Language Models (arXiv:2310.14451)

This project proposes a cascaded translation solution that does not require model retraining, solving the term accuracy problem in enterprise localization scenarios through multi-layer filtering and correction.

In enterprise localization scenarios, general machine translation has core contradictions:

1. Low Term Accuracy: Taking English-to-Arabic translation as an example, the general baseline term accuracy is only 36.67% (over 60% of professional vocabulary errors)
2. Disadvantages of Mixed Fine-tuning: Traditional mixed fine-tuning to inject domain terms leads to catastrophic forgetting—WMT 2023 tasks show that the BLEU score for English-to-Czech translation plummeted from 29.13 to 24.54, losing general language fluency.

The system adopts a four-tier cascaded pipeline (no model weight modification):

• Tier1 Translation Memory Caching: Hash table exact matching, O(1) time complexity, latency ~1ms, no API cost
• Tier2 Term Base Scanning: Regular expression (?i)\x08...\x08 for whole-word matching (case-insensitive) to avoid short abbreviation mismatches
• Tier3 MarianMT Local Inference: CPU inference with frozen weights (torch.no_grad()), generating a grammatically fluent first draft as a semantic anchor
• Tier4 Gemini 2.5 Post-editing: Inject term mapping table to replace general translations (e.g., correct "أداة الرصد" to "أداة مراقبة")

Core Components

• Python3.10+, HuggingFace Transformers, PyTorch, Google GenAI SDK, Pandas, XML ElementTree

Data Flow Output

1. Final translation: For Translation_Translated.xlsx
2. Translation memory update: clean_translation_memory.json
3. TBX standard term base: trados_enterprise_termbase.xml (supports SDL Trados import)
4. Real-time change log: Highlight term corrections

Security Isolation

• API keys are injected via environment variable GEMINI_API_KEY
• It is recommended to ignore local environment files in .gitignore to prevent sensitive information leakage

Metric	General Baseline	This Pipeline	Improvement Rate
Term Accuracy	36.67%	72.88%	+98.5%

The results are close to the best level of model retraining, with no risk of catastrophic forgetting and retaining general translation capabilities.

1. Resource-Constrained Environment: Low computing power input (only CPU runs MarianMT) + controllable API cost (call Gemini only when cache misses)
2. Translation Memory Enhancement: Extend from exact matching to a term-aware intelligent layer
3. Enterprise Term Governance: TBX standard export connects technology and localization toolchain (maintained by Python → used by Trados)

1. Latency Accumulation: Four-tier processing leads to higher end-to-end latency than pure local solutions
2. API Dependency: Gemini calls introduce network latency and cost; offline scenarios require degradation strategies
3. Term Base Maintenance: Accuracy depends on term base quality, requiring continuous input from domain experts

This project proves that through layered collaborative hybrid system design, even with limited resources, performance close to cutting-edge research can be achieved. It provides a referenceable engineering example for AI-assisted localization teams, emphasizing the pragmatic path of 'frozen baseline + cascaded correction'.