The 2nd Workshop on Automatic Simultaneous Translation

Challenges, Recent Advances, and Future Directions

Workshop proposal
Contact: autosimtrans.workshop@gmail.com or twitter.com/autosimtrans

1 Overview

Simultaneous translation, which performs translation concurrently with the source speech, is widely useful in many scenarios such as international conferences, negotiations, press releases, legal proceedings, and medicine. It combines the AI technologies of machine translation (MT), automatic speech recognition (ASR), and text-to-speech synthesis (TTS), is becoming a cutting-edge research field. As an emerging and interdisciplinary field, simultaneous translation faces many great challenges.

To promote the development in this field, we successfully held the first workshop on simultaneous translation at ACL 2020. We attracted 94 registered participants. We invited 6 keynote speakers, 4 from simultaneous translation, and 2 from human interpretation research. We also held a shared task on Chinese-English simultaneous translation and released an open dataset Baidu Speech Translation Corpus (BSTC) for open research, which covers speeches in a wide range of domains, including IT, economy, culture, biology, arts, etc. The shared task attracted 227 participants. However, affected by COVID-19, only a few submitted the results. The workshop also received 8 research paper submissions, with 5 accepted. The videos of invited talks were viewed 311 times (51.8 per talk). The videos for the 6 contributed talks were viewed 143 times (23.8 per talk). The peak zoom attendance was 25.

We have experiences of running a virtual workshop. All the talks and presentations are pre-recorded. We also run live Q&As for each session to let all the speakers and attendees for further discussion.

Since simultaneous translation is an active area in AI community, we propose to organize our second workshop in 2021. The workshop will bring together researchers and practitioners in machine translation, speech processing, and human interpretation, to discuss recent advances and open challenges of simultaneous translation. including:

  • Simultaneous translation paradigms: traditional pipeline (ASR-MT-TTS) or end-to-end (speech-speech); • Translation models: robust, high quality, and low latency;
  • Data resources: large and high-quality corpora for training simultaneous translation systems.
  • Evaluation methods: metrics to evaluate the translation quality and time latency;
  • Computer Aided Interpretation (CAI): improve the efficiency and quality of human interpreters;

2 Keynote Speakers

Following the tradition of our first workshop, we will have two sets of keynote speakers: 4 from simultaneous translation, and 2 from human interpretation research. We hope this workshop will greatly increase the communication and cross-fertilization between the two fields.

Simultaneous Translation Speakers:
  • Will Lewis, Principal Technical Project Manager, Microsoft Research, USA (confirmed)
  • Lucia Specia, Professor, Imperial College London, UK (confirmed)
  • Liang Huang, Assoc. Professor, Oregon State & Distinguished Scientist, Baidu Research, USA (confirmed)
  • Satoshi Nakamura, Professor, NAIST, Japan (tentatively yes)
Human Interpretation Speakers (both members of AIIC):
  • Hong Jiang, Sr. Lecturer, Dept. of Translation, Chinese University of Hong Kong, China (confirmed)
  • Laura Burian, Dean, Graduate School of Translation, Interpretation, and Language Education, and Professor of Chinese/English Translation and Interpretation, Middlebury Institute of International Studies (TBD)

We have 4 out of 6 speakers confirmed, and 3 out of 6 are female (Lucia, Hong, and Laura).

3 Diversity

As in our first workshop, we have made efforts to promote diversity in our invited speakers, organizers, and PC. 3 out of 6 invited speakers are female, so are 2 out of 6 organizers, and 7 out of 20 PC members.

4 Estimated Number of Attendees

Based on the fact that our first workshop (at ACL 2020) attracted 94 registrations, we estimate the number of partici- pants for our second workshop to be around 80–120.

5 Preferred Venues

We prefer our second workshop to appear at ACL 2021 or EMNLP 2021. But we have a special request. In 2020, both IWSLT and our workshop were held at ACL 2020, and there was a small overlap between the two workshops in terms of topics. After consultation with the organizers of IWSLT (Alex Waibel, Satoshi Nakamura, et al.) and participants of both workshops, we conclude that it is probably better to have these two workshops at separate conferences. Therefore, we’d appreciate if the ACL/EMNLP Workshop Chairs can take this into consideration. If IWSLT is going to be assigned to ACL 2021, please consider our workshop at EMNLP 2021, and vice versa.

6 Virtual Conference

We are well prepared if our second workshop has to be a virtual conference just like our first workshop at ACL 2020. In fact, we think it would be easier for the invited speakers when no travel is needed. The only major challenge for running a virtual workshop is time zone coordination, as we have participants from Asia, Europe, and the Americas. In 2020, we organized our first workshop into two sessions:

Session Pacific (PDT) Eastern (EDT) Europe (CET) East Asia (GMT+8) Speakers
#1 early morning
7:30–9:30am
late morning
10:30am–12:30pm
late afternoon
4:30-6:30pm
late evening / night
10:30pm-12:30am
Americas/Europe
#2 late afternoon
3-6:30pm
early evening
6-9:30pm
night time
12-3:30am
early morning
6-9:30am
Asia

As you can see, we tried our best to balance among the four time zones where most of our participants are from. The only suboptimal slot is Session #2 for European attendees, which was their night time. And we placed the talks into the slots that are most convenient for the speakers, e.g., we scheduled speakers from North America to Session #1 and those from Asia to Session #2. For our second workshop, we plan to have it based on this arrangement.

7 Shared Task

As we did in our first workshop, we plan to continue to organize the simultaneous translation shared tasks on Chinese- English and English-Spanish (or French). We will provide open datasets and evaluation environment. Participants are asked to submit their system description and show demonstration on the workshop. We will enlarge the corpus and extend the language pairs for our second workshop.

For more information, please see the Shared Task page.

8 Organizing Committee

  • Hua Wu, Baidu Inc., China, E-mail: wu_hua@baidu.com
  • Colin Cherry, Google, Canada, E-mail: colincherry@google.com
  • Liang Huang, Oregon State University and Baidu Research, USA, E-mail: liang.huang.sh@gmail.com
  • Zhongjun He, Baidu Inc., China, E-mail: hezhongjun@baidu.com
  • Qun Liu,Huawei Noah’s Ark Lab, Hong Kong, China, E-mail: qun.liu@huawei.com
  • Maha Elbayad, Université Grenoble Alpes, France, E-mail: maha.elbayad@infria.fr

The organizers come from companies and universities, having rich experiences in organizing conferences, workshops and evaluation campaigns. Their research interests include machine translation, natural language processing, machine learning, speech technologies, linguistics, etc.
· Hua Wu is the chief scientist of Baidu NLP and the chair of Baidu technical committee. Her primary research areas are machine translation and dialogue systems, and she has also worked on question answering, knowledge graph, and distributed representation. She served as program co-chair of ACL 2014, chaired international conferences such as ACL and IJCAI, and will be the program co-chair of AACL 2020.
· Colin Cherry is a research scientist at Google. He is currently serving as secretary of NAACL, and as an action editor of TACL. He received Best Paper Award at NAACL 2009. He co-organized two workshops on deep learning for low-resource languages: DeepLo 2018 (at ACL 2018) and DeepLo 2019 (at EMNLP 2019). He also served as research program co-chair for AMTA 2018.
· Liang Huang is an Associate Professor at Oregon State University and Distinguished Scientist of Baidu Research USA. He received a Best Paper Award at ACL 2008 and a Best Paper Honorable Mention at EMNLP 2016. He gave an invited talk at ACL 2019 on simultaneous translation.
· Zhongjun He is a Distinguished Architect of Baidu Inc. He leads Baidu machine translation team and has released several versions of Baidu’s simultaneous translation system since 2017. He organized the first simultaneous translation evaluation campaign in China in 2019.
· Qun Liu is Chief Scientist of Speech and Language Computing at Huawei’s Noak’s Ark Lab, Hong Kong, China. He was previously a Professor at Dublin City University and Chinese Academy of Sciences. He is widely recognized as a leading scientist in machine translation and natural language processing in general.
· Maha Elbayad recently completed her PhD from Universite ́ Grenoble Alpes. Her expertise includes neural machine translation, simultaneous translation, and more broadly on sequence-to-sequence prediction.

Here Hua, Colin, Liang, and Zhongjun also served on the Organizing Committee of the first workshop in 2020, and Qun gave an invited talk there. Two out of six members are female (Hua and Maha).

9 Steering Committee

  • Mark Liberman, Professor of Linguistics and Director of LDC, University of Pennsylvania, USA
  • Haifeng Wang, CTO, Baidu, China

10 Program Committee

(†: members from underrepresented demographic groups; 7 in total)

11 Sponsorships

We expect to have sponsorships from Baidu, Google, a few other companies, and some universities.

12 Awards

We will have a best paper award, a best presentation award, and a best poster award, sponsored by the above sponsors.