Fasoo/Sparrow Internship · Applied ML
A public reconstruction of an internship project that evaluated Korean-to-English product copy with multiple quality signals and post-edited only translations that required review.
Maximum observed during the internship evaluation
Share of selected translations whose grade improved after post-editing
Translation errors in multilingual product copy are uncommon, but their impact can be significant. A changed term, damaged placeholder, or altered technical identifier can affect both product behavior and user understanding. Reviewing every sentence at the same depth spends most of the effort on translations that are already acceptable.
The goal was not to remove final human review. It was to surface the translations most likely to contain errors. The pipeline needed to check meaning preservation, language quality, and format stability without a reference translation, then evaluate post-edited text with the same criteria.
LaBSE measures semantic similarity between Korean source text and English translation, while COMET provides a learned quality signal. A GEMBA-style LLM evaluation covers mistranslations that are difficult to express as rules. Deterministic checks inspect terminology, placeholders, technical identifiers, and unusual length changes.
The signals are normalized into a Q-Score, and only low-scoring translations are post-edited. Revised sentences pass through the same evaluation again, while final approval remains with a human reviewer. The public repository is a reduced reconstruction of the decision flow, not a copy of company code.
Semantic similarity alone can miss damaged placeholders, while deterministic rules cannot reliably identify fluent mistranslations. Each evaluator retains the error class it is best suited to detect instead of delegating every decision to one model.
Regenerating acceptable translations increases cost and regression risk. Only selected sentences are revised and evaluated again. The Q-Score prioritizes human review; it does not make the final approval decision.
Throughput and quality grades before and after post-editing were measured in the internship environment. The figures above describe the selected post-editing set, not all translations, and are not synthetic benchmark results from the public reconstruction.