Linear attention enhances inference efficiency of Transformer and has attracted research interests as an efficient backbone of language models. Existing linear attention based models usually exploit decay factor based positional encoding (PE), where attention scores decay exponentially with increasing relative distance. However, most work manually designs a non-trainable decay factor of exponential calculation, which limits further optimization. Our analysis reveals directly training decay factor is unstable because of large gradients. To address this, we propose a novel PE for linear attention named Disentangle to Decay (D2D). D2D disentangles decay factor into two parts to achieve further optimization and stable training. Moreover, D2D can be transformed into recurrent form for efficient inference. Experiments demonstrate that D2D achieves stable training of decay factor, and enhances performance of linear attention in both normal context length and length extrapolation scenarios.

Variational Autoencoder (VAE) is an effective framework to model the interdependency for non-autoregressive neural machine translation (NAT). One of the prominent VAE-based NAT frameworks, LaNMT, achieves great improvements to vanilla models, but still suffers from two main issues which lower down the translation quality: (1) mismatch between training and inference circumstances and (2) inadequacy of latent representations. In this work, we target on addressing these issues by proposing posterior consistency regularization. Specifically, we first perform stochastic data augmentation on the input samples to better adapt the model for inference circumstance, and then conduct consistency training on posterior latent variables to construct a more robust latent representations without any expansion on latent size. Experiments on En<->De and En<->Ro benchmarks confirm the effectiveness of our methods with about 1.5/0.7 and 0.8/0.3 BLEU points improvement to the baseline model with about 12.6× faster than autoregressive Transformer.