DiDi AI Labs Achieves Third Place in WMT2020 News Translation Task

DiDi AI Labs' NLP team recently secured the third place in the highly competitive news Chinese‑to‑English translation track of the WMT2020 competition, organized by the Association for Computational Linguistics (ACL). The competition attracted over forty teams from leading companies and universities worldwide.

The evaluation used the SacreBLEU metric, and DiDi's system achieved a BLEU score of 36.6, ranking third among all participants.

Technically, the team built upon a Transformer‑2 architecture, incorporating enhancements such as Self‑Attention, Relative Positional Attention, and a larger feed‑forward network. They employed iterative back‑translation and alternate knowledge distillation to generate high‑quality synthetic data, and applied data cleaning, selection, and model ensembling to boost translation quality. Additional techniques included domain transfer, topic mining with personalized weighting, as well as EDA and weight pruning to improve model robustness.

The resulting translation system has been deployed in various DiDi internationalization scenarios, including instant messaging translation, cross‑border dispute handling, and global operations. The team plans to continue advancing NLP and translation technologies, iterating on models and optimizing response speed for higher‑quality services.

References: 1. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation. 2. Ashish Vaswani et al. 2017. Attention is all you need. In Advances in Neural Information Processing Systems.