The AISP-SJTU Simultaneous Translation System for IWSLT 2022

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 208 - 215 May 26-27, 2022 c⃝2022 Association for Computational Linguistics The AISP-SJTU Simultaneous Translation System for IWSLT 2022 Qinpei Zhu1 Renshou Wu1 Guangfeng Liu1 Xinyu Zhu1 Xingyu Chen2 Yang Zhou1 Qingliang Miao1 Rui Wang2 Kai Yu1,2 1AI Speech Co., Ltd., Suzhou, China 2Shanghai Jiao Tong University, Shanghai, China Abstract This paper describes AISP-SJTU’s submis- sions for the IWSLT 2022 Simultaneous Trans- lation task. We participate in the text-to- text and speech-to-text simultaneous transla- tion from English to Mandarin Chinese. The training of the CAAT is improved by training across multiple values of right context win- dow size, which achieves good online perfor- mance without setting a prior right context window size for training. For speech-to-text task, the best model we submitted achieves 25.87, 26.21, 26.45 BLEU in low, medium and high regimes on tst-COMMON, correspond- ing to 27.94, 28.31, 28.43 BLEU in text-to-text task. 1 Introduction This paper describes the systems submitted by AI Speech Co., Ltd. (AISP) and Shanghai Jiaotong University (SJTU) for IWSLT 2022 Simultaneous Translation task. Two speech translation systems including cascade and end-to-end (E2E) for the Si- multaneous Speech Translation track, and a simul- taneous neural machine translation (MT) system for the text-to-text Simultaneous Translation track. The systems are focused on English to Mandarin Chinese language pair. For simultaneous speech translation, recent work tends to fall into two categories, cascaded systems and E2E systems. And the cascaded system of- ten outperforms the fully E2E approach. Only one work (Ansari et al., 2020; Anastasopoulos et al., 2021) shows that the E2E model can achieve better results than the cascaded model. In their work they introduce pre-training (Stoian et al., 2020; Dong et al., 2021; Wang et al., 2020b) and data augmen- tation techniques (Pino et al., 2020; Xu et al., 2021) to E2E models. Therefore, in this paper, we hope to optimize the speech translation model from two aspects. First, we aim to build a robust cascade model and learn best practices from WMT evalua- tion activities (Wu et al., 2020; Meng et al., 2020; Zeng et al., 2021), such as back translation (Sen- nrich et al., 2015; Edunov et al., 2018; Lample et al., 2017). Second, we explore various self- supervised learning methods and introduce as much semi-supervised data as possible towards finding the best practice of training cascaded speech-to- text (S2T) models. In our settings, ASR data, MT data, and monolingual text data are all considered in a progressively training framework. We only trained one E2E model, and its BLEU is 22.49 with 1272 AL. Due to the huge difference in the scale of training data from the cascaded model, E2E per- formance is far lower than that of the latter. The cascaded S2T final performance on the MuST-C V2 test set is 25.87, 26.21, 26.45 BLEU with low, medium and high regimes. In addition, we also participate in the simultane- ous text-to-text (T2T) task. Our system is based on an …

Content truncated to 500 words