The Xiaomi Text-to-Text Simultaneous Speech Translation System for IWSLT 2022

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 216 - 224 May 26-27, 2022 c⃝2022 Association for Computational Linguistics The Xiaomi Text-to-Text Simultaneous Speech Translation System for IWSLT 2022 Bao Guo1∗ Mengge Liu2∗† Wen Zhang1 Hexuan Chen1 Chang Mu1 Xiang Li1 Jianwei Cui1 Bin Wang1 Yuhang Guo2 1Xiaomi AI Lab, Beijing, China 2Beijing Institute of Technology, Beijing, China guobao@xiaomi.com 3120201046@bit.edu.cn {zhangwen17,chenhexuan,muchang1,lixiang21}@xiaomi.com {cuijianwei,wangbin11}@xiaomi.com guoyuhang@bit.edu.cn Abstract This system paper describes the Xiaomi Trans- lation System for the IWSLT 2022 Simul- taneous Speech Translation (noted as SST) shared task. We participate in the English- to-Mandarin Chinese Text-to-Text (noted as T2T) track. Our system is built based on the Transformer model with novel techniques bor- rowed from our recent research work. For the data filtering, language-model-based and rule-based methods are conducted to filter the data to obtain high-quality bilingual parallel corpora. We also strengthen our system with some dominating techniques related to data augmentation, such as knowledge distillation, tagged back-translation, and iterative back- translation. We also incorporate novel training techniques such as R-drop, deep model, and large batch training which have been shown to be beneficial to the naive Transformer model. In the SST scenario, several variations of wait-k strategies are explored. Furthermore, in terms of robustness, both data-based and model-based ways are used to reduce the sen- sitivity of our system to Automatic Speech Recognition (ASR) outputs. We finally de- sign some inference algorithms and use the adaptive-ensemble method based on multiple model variants to further improve the perfor- mance of the system. Compared with strong baselines, fusing all techniques can improve our system by 2~3 BLEU scores under differ- ent latency regimes. 1 Introduction In the IWSLT 2022 Evaluation Campaign, our team at Xiaomi AI Lab participates in one Simultane- ous Speech Translation task (Anastasopoulos et al., 2022), which is the Text-to-Text track in English to Mandarin Chinese translation direction. We first introduce the techniques used in our final submitted ∗Equal contribution. † The work was done during the author’s internship at Xiaomi. system from four aspects: data, model, inference, and robustness. Data-related techniques are introduced from two perspectives: data augmentation and domain- related data selection. For data augmentation, we employ technologies such as back-translation (BT) (Sennrich et al., 2016a), knowledge distil- lation (KD) (Kim and Rush, 2016), and iterative back-translation (Hoang et al., 2018) etc. to gener- ate large-scale synthetic bilingual datasets, which have been proved to be very effective in the field of machine translation. We also use the technology of Tagged Back-Translation (TaggedBT) (Caswell et al., 2019), that is, prepending a reserved token <BT> to the beginning of the synthetic source sen- tence in the training set, so that the model could distinguish the originality of the source sentence. Meanwhile, the effects of different combinations of multiple training sets on the model performance are explored. For domain-related data selection, differences in the domains of the training and test sets can have a large negative impact on the re- sults on the test sets. To …

Content truncated to 500 words