Machine translation (MT) systems often struggle with gender when translating from a gender-neutral language like English to a gendered language such as Romanian, frequently defaulting to masculine forms or reinforcing stereotypes. A new research paper introduces a hybrid pipeline to address this issue.
The proposed system uses a fine-tuned large language model (LLM) to detect the intended gender of target words in English sentences and inserts inline gender hint tags. These tagged sentences are then processed by a Transformer model fine-tuned to generate morphologically correct Romanian translations.
To support this approach, the authors introduce three novel datasets for gender disambiguation and translation. Their method improves gender accuracy on the WinoMT and WinoGender benchmarks by over 40 percentage points compared to a baseline MT system.
This work is the first to explicitly address and evaluate gender bias in English-Romanian MT using both LLM inference and tag-aware translation, marking a significant step toward more equitable language technology.