Uncertainty Quantification for Evaluating Machine Translation Bias

In machine translation (MT), when the source sentence includes a lexeme whose gender is not overtly marked, but whose target-language equivalent requires gender specification, the model must infer the appropriate gender from the context and/or external knowledge. Studies have shown that MT models exhibit biased behaviour, relying on stereotypes even when they clash with contextual information. We posit that apart from confidently translating using the correct gender when it is evident from the input, models should also maintain uncertainty about the gender when it is ambiguous. Using recently proposed metrics of semantic uncertainty, we find that models with high translation and gender accuracy on unambiguous instances do not necessarily exhibit the expected level of uncertainty in ambiguous ones. Similarly, debiasing has independent effects on ambiguous and unambiguous translation instances.