How Tencent Dominated WMT2021: Winning Five News‑Track Translation Tasks

WMT2021 International Machine Translation Competition Results

The WMT2021 competition, organized by the Association for Computational Linguistics (ACL), released the final rankings for its news‑track tasks. Tencent secured first place in five translation tasks: Chinese‑English bidirectional, Japanese‑English bidirectional, and English‑German limited‑resource translation.

The winning teams were composed of members from TencentMT Oteam’s WeChat translation team, Tencent AI Lab, and the "translation‑jun" group.

Since its inception in 2006, WMT has become the most authoritative benchmark for evaluating machine‑translation systems, attracting top teams from Microsoft, Facebook, DeepMind, Huawei, Baidu, Kingsoft, NICT, Tsinghua University, Shanghai Jiao Tong University, and many others.

Submissions were evaluated using the BLEU automatic metric, which measures the similarity between system outputs and reference translations. A total of 82 teams submitted 1,308 valid runs, indicating intense competition.

After multiple rounds of evaluation, Tencent’s systems topped the automatic metrics in the four Chinese‑English and Japanese‑English tasks, surpassing all systems that used either limited resources or external (unrestricted) resources.

In the English‑German direction, Tencent placed second, beating every limited‑resource system, while the first‑place system was a non‑restricted system that leveraged external data.

Below are the top‑10 ranking tables for each language pair:

TencentMT Oteam originated from Tencent’s internal push for open‑source collaboration. Established in early 2019 under the Tencent Technical Committee, it unites multiple translation groups, centralizes resources, and deeply optimizes core MT modules to build a powerful translation engine and services.

Within Tencent, machine‑translation technology is widely deployed in products such as WeChat, QQ, the browser, email, and auxiliary translation tools.

—END—