Can Explicit Gender Information Improve Zero-Shot Machine Translation?

AbstractLarge language models (LLMs) have demonstrated strong zero-shot machine translation (MT) performance but often exhibit gender bias that is present in their training data, especially when translating into grammatically gendered languages. In this paper, we investigate whether explicitly providing gender information can mitigate this issue and improve translation quality. We propose a two-step approach: (1) inferring entity gender from context, and (2) incorporating this information into prompts using either Structured Tagging or Natural Language. Experiments with five LLMs across four language pairs show that explicit gender cues consistently reduce gender errors, with structured tagging yielding the largest gains. Our results highlight prompt-level gender disambiguation as a simple yet effective strategy for more accurate and fair zero-shot MT.