2019 (modified: 11 Nov 2022)ICML 2019Readers: Everyone
Abstract:We introduce a new beam search decoder that is fully differentiable, making it possible to optimize at training time through the inference procedure. Our decoder allows us to combine models which o...