Hybrid Pipeline Reduces Gender Bias in Romanian Translation

A new translation pipeline mitigates gender bias in English-to-Romanian tasks by combining LLM-based gender classification with fine-tuned NMT. By inserting inline gender hint tags into source sentences, the system improves WinoMT and WinoGender benchmark accuracy by over 40 percentage points.