Mitigating Gender Bias in English to Romanian Machine Translation
Ioana Grigore, Sergiu Nisioi
cs.CL, cs.AI
2026-08-09
A fine-tuned 1B LLaMA tags gender in English sentences and a tag-aware Transformer renders correct Romanian, lifting WinoMT gender accuracy from 59% to 96%.
English barely marks gender on occupations; Romanian requires agreement across nouns, adjectives and participles. When an MT system translates "The doctor said she...", it holds no gender information, defaults to masculine, and replicates stereotypes along the way: doctors come out masculine, nurses feminine. On WinoMT-style benchmarks, anti-stereotypical gender accuracy sits near coin flip.
A two-stage pipeline from a University of Bucharest team: decide gender first, then make it visible to the translator.
Three datasets are released: D1 (11,472 single-entity sentences), D2 (996 two-entity sentences with each target word appearing in all three gender variants), and EnRoGend (1,974 parallel English-Romanian minimal pairs over 82 occupations, masculine and feminine balanced, split pair-locked against leakage).
| System | WinoMT Pro | WinoMT Anti | WinoGender |
| Raw Transformer | 59.05% | 50.13% | 50.28% |
| This pipeline | 96.34% | 93.68% | 91.53% |
| GPT-5.2 zero-shot | 79.23% | 58.79% | 67.71% |
Gains run 37 to 43 points across the three benchmarks, and the pro/anti stereotype gap narrows from 8.92 to 2.66 points. The GPT-5.2 row is the cautionary one: 79.23% on pro-stereotypical items but 58.79% on anti-stereotypical, so the zero-shot frontier model carries a heavier stereotype pull than the small baseline. On validation, the fully fine-tuned system reaches BLEU 97.29 and COMET 0.853; LoRA on the translator underperforms badly (COMET 0.65 to 0.73), suggesting tag-conditioned morphological control needs more adaptation capacity than parameter-efficient tuning offers.
One detail that matters: the benchmark test sets have no Romanian reference translations, so gender accuracy was judged by a native speaker with spot checks, not by an automatic metric.
This is a cheap debiasing recipe: a quantized 1B classifier trains on a single L4 GPU, and the translator is an off-the-shelf opus-mt model. For any gender-neutral to gendered direction (English to Hungarian, Turkish, Finnish, Persian), inline tags cost far less than curating balanced training corpora. English-Spanish and English-French had prior work; this is the first method to explicitly address and evaluate gender bias in English-Romanian MT using LLM inference plus tag-aware translation.
One language pair only. The benchmark evaluation rests on a single native speaker with spot checks; sample sizes and inter-annotator agreement are not reported. Translations still lose diacritics and pick wrong words (the authors note both raw and fine-tuned systems do), and fluency beyond gender correctness is not systematically evaluated. The pipeline adds a classification forward pass over direct MT. D1, D2 and EnRoGend are synthetic corpora, cleaner than real text, and how much of the gain survives on natural data is untested.