1 result for “transformer machine translation”

The Impact of Transformer Architectures on Machine Translation

Transformer architectures displaced recurrent models as the dominant paradigm in neural machine translation within three years of their introduction. This paper traces that displacement, arguing that the decisive advantage was not raw accuracy but parallelisability during training, which converted translation quality into a problem of available compute. It evaluates the empirical record across high- and low-resource language pairs and considers where the architecture's assumptions still break down.