HW-TSC at IWSLT2023: Break the Quality Ceiling of Offline Track via Pre-Training and Domain Adaptation

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 187–193 July 13-14, 2023 c⃝2023 Association for Computational Linguistics HW-TSC at IWSLT2023: Break the Quality Ceiling of Offline Track via Pre-Training and Domain Adaptation Zongyao Li, Zhanglin Wu, Zhiqiang Rao, Xie YuHao, Guo JiaXin, Daimeng Wei, Hengchao Shang, Wang Minghan, Xiaoyu Chen Zhengzhe YU, Li ShaoJun, Lei LiZhi, Hao Yang Huawei Translation Service Center, Beijing, China {lizongyao,wuzhanglin2,raozhiqiang,xieyuhao2,guojiaxin1, weidaimeng,shanghengchao,wangminghan,chenxiaoyu35, yuzhengzhe,lishaojun18,leilizhi,yanghao30}@huawei.com Abstract This paper describes HW-TSC’s submissions to the IWSLT 2023 Offline Speech Transla- tion task, including speech translation of talks from English to German, English to Chinese and English to Japanese. We participated in all three tracks (Constrained training, Constrained with Large Language Models training, Uncon- strained training), with using cascaded architec- tures models. We use data enhancement, pre- training models and other means to improve the quality of ASR, and use a variety of tech- niques including R-Drop, deep model, domain data selection, etc. to improve the quality of NMT. Compared with last year’s best results, we have improved by 2.1 BLEU in the MuST-C English-German test set. 1 Introduction The goal of the Offline Speech Translation Task is to examine automatic methods for translating audio speech in one language into text in the tar- get language. In recent years, end-to-end system and cascade system are fundamental pipelines for speech translation tasks. Traditional cascade sys- tem is comprised of continuing parts, automatic speech recognition (ASR) is responsible for gener- ating transcripts from audios and machine transla- tion (MT) model aims at translating ASR outputs from source language into target language. ASR model like Conformer (Gulati et al., 2020) and S2T- Transformer (Synnaeve et al., 2019) are commonly used. MT models like Transformer (Vaswani et al., 2017) can be considered as a standard configura- tion. The End-to-end systems use a model to di- rectly recognize speech into target text in another language. The cascade system will cause some "missing information" due to the two encoding and decoding processes of ASR and MT. At the same time, the disadvantage of the end-to-end system is the lack of sufficient training data. However, with a fully trained cascade system, the accuracy of ASR and MT will reach a higher level. So from the results, the BLEU of the cascaded system will be higher than that of the end-to-end system. Currently in the industry, the mainstream speech translation sys- tem is still based on the cascade system. We use the cascade system for this task, mainly to further improve the performance of speech translation. In this work, we carefully filter and preprocess the data, and adopt various enhancement tech- niques, such as pre-training model, data enhance- ment, domain adaptation, etc., to optimize the performance of ASR. We build machine transla- tion systems with techniques like back translation (Edunov et al., 2018), domain adaptation and R- drop (Wu et al., 2021), which have been proved to be effective practices. The main contribution of this paper can be sum- marized as follows: 1) According …

Content truncated to 500 words