Traditional multilingual neural machine translation (MNMT) uses a single
model to translate all directions. However, with the increasing scale of
language pairs, simply using a single model for massive MNMT brings new
challenges: parameter tension and large computations. In this paper, we revisit
multi-way structures by assigning an individual branch for each language
(group). Despite being a simple architecture, it is challenging to train
de-centralized models due to the lack of constraints to align representations
from all languages. We propose a localized training recipe to map different
branches into a unified space, resulting in an efficient detachable model,
Lego-MT. For a fair comparison, we collect data from OPUS and build the first
large-scale open-source translation benchmark covering 7 language-centric data,
each containing 445 language pairs. Experiments show that Lego-MT (1.2B) brings
gains of more than 4 BLEU while outperforming M2M-100 (12B) (We will public all
training data, models, and checkpoints)

本文提出了一种名为 Lego-MT 的多向分支神经机器翻译模型，通过局部化训练，将不同的分支映射到一个统一的空间中，达到高效、可分离的效果，并在收集的包含 7 个语言中心数据的大型开源翻译基准测试中，取得了超过 4 个 BLEU 的提升，优于 M2M-100 模型。