A new study by researchers from Google and Imperial College London challenges a core assumption in AI translation evaluation: that a single metric can capture both semantic accuracy and naturalness of ...