Abstract:

In Alibaba's ecological environment, we need to satisfy the translation demands of overseas e-commerce platforms of tens of millions or even hundreds of millions of levels every day. The performances of machine translations are traditionally evaluated by the metric BLEU when the golden references are provided. However, in the case of model inference or production deployment, golden references are usually expensively available. To address the issue of translation quality estimation without reference, Alibaba proposes a general framework for automatic evaluation of the machine translation (MT) output in both sentence and word levels, a novel algorithm based on bidirectional self-attention mechanism which greatly improves Pearson correlation coefficient between automatic and manual evaluations. This solution can not only serve in a variety of computer-aided applications such as translation result screening before release but also support to optimize MT engines via an error feedback system in multiple business scenarios including automatic speech, communication and e-commerce translations.

Bio:

Jiayi Wang is working as a Senior Algorithm Engineer in the machine translation team of the Machine Intelligence Technology Lab at Damo Academy, Alibaba, responsible for developing real-time automatic quality inspection systems for online machine translation and speech recognition (ASR) engines. She received her Master Degree in the Department of Applied Mathematics and Statistics, Johns Hopkins University in 2015. Prior to Alibaba, she worked as a staff researcher in Social Science Research Institute (SSRI) at Duke University in the field of computational genetics from 2016 to 2017.