Edinburgh’s End-to-End Multilingual Speech Translation System for IWSLT 2021
20718Authors
nan
Abstract
AbstractThis paper describes Edinburgh’s submissions to the IWSLT2021 multilingual speech translation (ST) task. We aim at improving multilingual translation and zero-shot performance in the constrained setting (without using any extra training data) through methods that encourage transfer learning and larger capacity modeling with advanced neural components. We build our end-to-end multilingual ST model based on Transformer, integrating techniques including adaptive speech feature selection, language-specific modeling, multi-task learning, deep and big Transformer, sparsified linear attention and root mean square layer normalization. We adopt data augmentation using machine translation models for ST which converts the zero-shot problem into a zero-resource one. Experimental results show that these methods deliver substantial improvements, surpassing the official baseline by > 15 average BLEU and outperforming our cascading system by > 2 average BLEU. Our final submission achieves competitive performance (runner up).
Keywords
Full Content
Proceedings of the 18th International Conference on Spoken Language Translation, pages 1–29 Bangkok, Thailand (Online), August 5–6, 2021. ©2021 Association for Computational Linguistics 1 FINDINGS OF THE IWSLT 2021 EVALUATION CAMPAIGN Antonios Anastasopoulos George Mason U. Ondˇrej Bojar Charles University Jacob Bremerman UMD Roldano Cattoni FBK Maha Elbayad Facebook Marcello Federico Amazon AI Xutai Ma JHU/Facebook Satoshi Nakamura NAIST Matteo Negri FBK Jan Niehues Maastricht U. Juan Pino Facebook Elizabeth Salesky JHU Sebastian St¨uker KIT Katsuhito Sudoh NAIST Marco Turchi FBK Alex Waibel CMU/KIT Changhan Wang Facebook Matthew Wiesner JHU Abstract The evaluation campaign of the International Conference on Spoken Language Translation (IWSLT 2021) featured this year four shared tasks: (i) Simultaneous speech translation, (ii) Offline speech translation, (iii) Multilingual speech translation, (iv) Low-resource speech translation. A total of 22 teams participated in at least one of the tasks. This paper de- scribes each shared task, data and evaluation metrics, and reports results of the received sub- missions. 1 Introduction The International Conference on Spoken Lan- guage Translation (IWSLT) is the premier an- nual scientific conference for all aspects of spo- ken language translation. For 18 years running (Akiba et al., 2004; Eck and Hori, 2005; Paul, 2006; Fordyce, 2007; Paul, 2008, 2009; Paul et al., 2010; Federico et al., 2011, 2012; Cettolo et al., 2013, 2014, 2015, 2016, 2017; Niehues et al., 2018, 2019; Ansari et al., 2020), the conference or- ganizes and sponsors open evaluation campaigns around key challenges in simultaneous and con- secutive translation, under real-time/low latency or offline conditions and under low-resource or multilingual constraints. System descriptions and results from participants’ systems and scientific papers related to key algorithmic advances and best practice are published in proceedings and presented at the conference. IWSLT is also the venue of the SIGSLT, the Special Interest Group on Spoken Language Translation of ACL, ISCA and ELRA. With its long track record, IWSLT benchmarks and proceedings serve as reference for all researchers and practitioners working on speech translation and related fields. This paper reports on the evaluation campaign organized by IWSLT 2021, which features four shared tasks: • Simultaneous speech translation, address- ing low latency translation of talks, from En- glish to German and English to Japanese, ei- ther from a speech file into text, or from a ground-truth transcript into text; • Offline speech translation, proposing speech translation of talks from English into German, using either cascade architectures or end-to-end models, able to directly translate source speech into target text; • Multilingual speech translation, focusing 2 Team Organization APPTEK AppTek, Germany (Bahar et al., 2021b) BUT Brno University of Technology, Czech Republic (Vydana et al., 2021) ESPNET-ST ESPnet-ST group, Johns Hopkins University, USA (Inaguma et al., 2021) FBK Fondazione Bruno Kessler, Italy (Papi et al., 2021) FAIR FAIR Speech Translation (Tang et al., 2021a) HWN Huawei Noah’s Ark Lab, China (Zeng et al., 2021) HW-TSC Huawei Translation Services Center, China IMS University of Stuttgart, Germany (Denisov et al., 2021) KIT Karlsruhe Institute of Technology, Germany (Nguyen et al., 2021; …
Content truncated to 500 words
Labeled Text
AbstractThis paper describes Edinburgh’s submissions to the IWSLT2021 multilingual speech translation (ST) task. We aim at improving multilingual translation and zero-shot performance in the constrained setting (without using any extra training data) through methods that encourage transfer learning and larger capacity modeling with advanced neural components. We build our end-to-end multilingual ST model based on Transformer, integrating techniques including adaptive speech feature selection, language-specific modeling, multi-task learning, deep and big Transformer, sparsified linear attention and root mean square layer normalization. We adopt data augmentation using machine translation models for ST which converts the zero-shot problem into a zero-resource one. Experimental results show that these methods deliver substantial improvements, surpassing the official baseline by > 15 average BLEU and outperforming our cascading system by > 2 average BLEU. Our final submission achieves competitive performance (runner up).
Publication Details
- Paper ID:
- 20718
- Year:
- 2021
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan