# Neural Machine Translation with Byte-Level Subwords

Changhan Wang, Kyunghyun Cho and Jiatao Gu

The Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI)., 2020

Abstract
Almost all existing machine translation models are built on top of character-based vocabularies: characters, subwords or words. However, characters (and words) are usually long-tail distributed, especially for noisy texts or character-rich languages like Japanese and Chinese, which inflates the vocabulary with rare symbols. In pursuit of extreme compactness, byte vocabulary was explored, but its high computational cost becomes a blocker in practice. In this paper, we investigate byte-level subwords, specifically byte-level BPE (BBPE), which is similarly compact with character set and generic without out-of-vocabulary tokens, but is more efficient than pure bytes. And we claim that contextualizing BBPE embeddings is beneficial, which can be simply implemented by 1-layer convolution or GRU. Our empirical results show that BBPE can perform better than char-level BPE with less than 25% of the volume. In the multilingual setting, BBPE enlarges vocabulary sharing among different languages and achieves better performance in universal many-to-En models.

[paper] [code]

@article{wang2019neural,