Non-Autoregressive Translation by Learning Target Categorical Codes
This addresses the problem of lower translation quality in non-autoregressive models for machine translation, representing an incremental improvement.
The paper tackles the accuracy gap in non-autoregressive translation models by proposing CNAT, which learns categorical codes to model dependencies, achieving comparable or better performance in machine translation tasks.
Non-autoregressive Transformer is a promising text generation model. However, current non-autoregressive models still fall behind their autoregressive counterparts in translation quality. We attribute this accuracy gap to the lack of dependency modeling among decoder inputs. In this paper, we propose CNAT, which learns implicitly categorical codes as latent variables into the non-autoregressive decoding. The interaction among these categorical codes remedies the missing dependencies and improves the model capacity. Experiment results show that our model achieves comparable or better performance in machine translation tasks, compared with several strong baselines.