2000 character limit reached
Semi-Autoregressive Training Improves Mask-Predict Decoding (2001.08785v1)
Published 23 Jan 2020 in cs.CL, cs.LG, and stat.ML
Abstract: The recently proposed mask-predict decoding algorithm has narrowed the performance gap between semi-autoregressive machine translation models and the traditional left-to-right approach. We introduce a new training method for conditional masked LLMs, SMART, which mimics the semi-autoregressive behavior of mask-predict, producing training examples that contain model predictions as part of their inputs. Models trained with SMART produce higher-quality translations when using mask-predict decoding, effectively closing the remaining performance gap with fully autoregressive models.