Quality Estimation with Transformer and RNN Architectures

Yulin Zhang, Chong Feng, Hongzheng Li

Abstract: The goal of China Conference on Machine Translation (CCMT 2019) Shared Task on Quality Estimation (QE) is to investigate automatic methods for estimating the quality of $$ {\text{Chinese}}\!\leftrightarrow\!{\text{English}} $$ machine translation results without reference translations. This paper presents the submissions of our team for the sentence-level Quality Estimation shared task of CCMT19. Considering the good performance of neural models in previous shared tasks of WMT, our submissions also include two neural-based models: one is Bi-Transformer which proposes the model as a feature extractor with a bidirectional transformer and then processes the semantic representations of source and the translation output with a Bi-LSTM predictive model for automatic quality estimation, and the other BiRNN architecture uses only two bi-directional RNNs (bi-RNN) with Gated Recurrent Units (GRUs) as encoders, and learns representation of the source and translation sentence pairs to predict the quality of translation outputs.