% grep -r "back-translation" /posts/
Back-translation
How Native Is Eiken's Official Website English?
Back-translation scores tend to favor literal translations. So what does "high-scoring but not quite natural" English look like? I looked at the English on Eiken's official website and thought about the structural reasons behind literal-sounding translations.
My Translation Scored 0.96, But the English Version Fell Apart
A high back-translation score doesn't guarantee native-level English. I ran into a case where a 0.96-scoring translation was completely broken for native readers — and learned that translation quality has two orthogonal axes.
Testing 3 Different Translation Prompt Patterns
I tweaked the translation prompts to get more natural output — and accidentally polluted my own evaluation mechanism. Here's what I found.
AI Translation Model Comparison: Back-Translation Scores Across the Gemini Family
Comparing back-translation scores across multiple AI models including Gemini and Claude. Looking at how translation accuracy varies by article content, embedding model characteristics, and the difference between translate and evaluate modes.
Automated Translation Quality Verification with AI
I built a system that automatically verifies translation quality using back-translation — no reference translations needed. Scores are calculated with Gemini Embedding + pgvector cosine similarity, comparing results across multiple models.