Alibaba's Advances in Multilingual Neural Machine Translation: Research and Practice

Alibaba aims to eliminate language barriers for its global commerce platforms by supporting translation for more than 200 languages, which entails handling over 40,000 language pairs and presents significant data‑collection and model‑capacity challenges.

Multilingual neural machine translation (NMT) offers two main benefits: reduced deployment and training costs through a single model, and knowledge sharing that improves low‑resource language pairs, but it also introduces challenges such as the need for high‑quality multilingual data and sufficient model capacity.

Alibaba evaluated two mainstream frameworks: a separated encoder‑decoder per language pair, which scales poorly, and a universal shared encoder‑decoder model, which suffers from capacity limits and language conflicts. To balance universality and specificity, Alibaba introduced an intermediate‑language module placed between a shared encoder and decoder, enforcing reconstruction and semantic consistency losses to learn a language‑agnostic representation.

Experiments on WMT and internal test sets showed that the intermediate‑language model matches the performance of individually trained models and outperforms the universal model, achieving comparable BLEU scores while providing better zero‑shot translation capabilities.

To further improve zero‑shot performance, Alibaba developed a repair‑translation pipeline: pseudo‑parallel data generated by back‑translation is refined by a multilingual repair model (DR) trained on triplets of source, noisy translation, and repaired translation, and an iterative process alternates between repairing data and retraining the NMT model.

Alibaba also explored integrating large pre‑trained language models (BERT, MASS, mBART) into multilingual NMT. A dynamic fusion approach uses attention to select useful layers from the pre‑trained model, while a lightweight adapter method inserts small feed‑forward modules into BERT layers, fixing the pre‑trained parameters to avoid catastrophic forgetting and to keep training stable.

To meet production latency requirements, Alibaba applied engineering optimizations: a deep encoder with a shallow decoder, shared attention weights across decoder layers, and a shortlist vocabulary prediction, collectively achieving a 3–4× speedup in decoding.

The resulting multilingual NMT system has been deployed on Alibaba platforms such as AliExpress and Alibaba Cloud, providing real‑time translation for over 200 language pairs and supporting the company’s global buying, selling, travel, and payment initiatives.