Lossless Speedup of Autoregressive Translation with Generalized Aggressive Decoding

Different from previous work accelerating translation at the cost of quality loss, we propose Generalized Aggressive Decoding (GAD) -- a novel decoding paradigm for lossless speedup of autoregressive translation, through the collaboration of autoregressive and non-autoregressive translation (NAT) of the Transformer. At each decoding iteration, GAD aggressively decodes a number of tokens with NAT as a draft and then verifies them in the autoregressive manner, where only the tokens that pass the verification are accepted as decoded tokens. GAD can achieve the same results as autoregressive translation but much more efficiently because both NAT drafting and autoregressive verification compute in parallel. We conduct experiments in four standard WMT benchmarks and confirm that the vanilla GAD yields exactly the same results as greedy decoding with an around $3\times$ speedup, and that its variant (GAD++) with an advanced verification strategy not only outperforms the greedy translation and even achieves the comparable translation quality with the beam search result, but also further improves the decoding speed, resulting in an around $5\times$ speedup over autoregressive translation. Moreover, GAD can be easily generalized for lossless speedup of other seq2seq tasks like Abstractive Summarization, and benefit more from stronger computing devices, demonstrating its potential to become a de facto decoding paradigm in the future. Our models and codes are available at https://github.com/hemingkx/GAD.