Earlier this month we had the pleasure of participating in the annual conference (virtual this year) of the Association for Machine Translation in the Americas (AMTA). AMTA was founded in 1991, along with sister groups in Asia-Pac (AAMT) and Europe (EAMT) as a part of the International Association for Machine Translation.
This was my first attendance at AMTA, after years of focusing mainly on TAUS as my go-to for the latest on machine translation and other language technology developments. I had always thought of AMTA as purely an R&D event, but this year President Steve Richardson and committee chairs Janice Campbell, Natalia Levitina and Ray Flournoy, made a very effective and concerted effort to communicate to us ahead of time that, in fact, AMTA attendance and topics are roughly evenly divided in thirds, between R&D/academia, Commercial and Government. SDL signed up as a sponsor; had a virtual “booth;” our inimitable Bart Maczynski did a talk in the commercial track, and I had the pleasure of co-moderating a couple of presentations, as well as attending as many of the 65+ sessions (workshops, tutorials, keynotes, sessions, demos…) as humanly possible.
First, kudos to the event organizers for making it work in Microsoft Teams. With 410 registered attendees from around the world, this was not an easy challenge. Each session was its own Teams meeting (recorded), and virtual booths were set up as channels.
There were tutorials and workshops on Oct 6, and then the main 3 days (Oct 7-9) had 6 keynotes, interspersed among many sessions and divided into 16 research, 22 commercial and 8 government tracks. I attended the commercial and government sessions, but besides dividing them that way, in retrospect I think you can sort the themes of the sessions into Tools and People sessions, the people equating mostly to post-editors.
If we start with what I consider the “People” sessions, it was clear just how far MTPE has evolved. Many of the sessions served to solidify and provide quantitative measures for what we know already in practice. But seeing multiple sessions in the track enabled participants to observe progress in the attitudes toward, and expertise in, Post-Editing, for the industry overall.
Some notable sessions were:
Session | Topic | Take-away |
W1 | Best practices for Machine Translation Post-Editor training | In the speaker’s job, most of the time MT replaces the T in TEP and so should also be followed by edit and proof steps, to attain human quality. Cool photo of a minimalist PE workstation! |
W1 | Greek experiment with measuring Post- Editor time savings | A point of debate was that the study found PE less cognitively demanding than translating from scratch, but this contrasted with audience experience for technical content, where the close attention needed for corrections was perceived as more demanding. |
W1 | Panel on future of Machine Translation Post-Editing | TM is still needed for the comfort of translators, that MT parity with human translation is not expected for some time to come, and that post-editors will thus still be needed. |
C12 | U of Maryland experiments with MT productivity | In general, students found that editing was clearer and faster with TM fuzzies versus PE. Referenced SDL blogpost on “Edit Distance: Not a Miracle Cure.” |
C18 | Understanding Obstacles to Enterprise MT adoption | Bart showed some nice spiderweb diagrams for different use cases, like chatbots and eDiscovery, to map against quantitative attributes like speed, security, quality. |
If we turn our attention to what I would label as the “Tools” sessions, the first observation is that there are quite a few more “Tools” sessions vs “People” sessions, and that’s not even counting those in the Research track. And this is one of the things I loved about this AMTA conference – it was an amazing snapshot of virtually everything happening on the machine translation technology front…the things that can be talked about, that is!
Some highlights:
Session | Topic | Take-away |
W1 | Automated PE support tool | Online tool allowing users to submit a text and get a machine-learning-based prediction of the post-editing effort that will be needed. Interesting in that it shows how machines can help – not replace – human translation. |
W1 | Multimodal tools for PE | Going beyond keyboard and mouse, to handwriting, voice, touchscreen, eye tracking! |
K2 | “Faithfulness” as a proposed new metric in Natural Language Generation | Translation may be high in fluency and adequacy, yet not “faithful;” that is, achieving pragmatics equivalence. |
K6 | Latest Google Research on Transcribe Mode in Google Translate | Team explored translating a sentence during transcription, while minimizing latency and maximizing quality. One option was to allow publishing of a pre-version with MT that would be revised later with more context. Future focus areas will include more on Speech to Speech. |
G6 | American sign language to English MT app | ASL Signer to English: video device captures the ASL signs, signs are processed by Dragonfly which translates to English and displays the text. For English speakers to ASL, the spoken voice is captured by Dragonfly and converted to text, and then converted to ASL, and text plus signing avatar are displayed. Lots of complexity in sign processing due to variability of signers and other factors. |
When I was debriefing afterward on the overall conference with my colleague Bart Maczynski, who presented “Understanding Challenges to Enterprise Machine Translation Adoption,” he remarked on how the MT use cases he observed during the conference were mostly related to outbound translation workflow: using MT for productivity gains in an otherwise regular localization workflow. This contrasts with our SDL experience which shows that the overall volumes going through MT in a localization workflow are just a fraction of the huge volumes of content in those MT use cases that involve external content coming in, for example eDiscovery, market surveillance, intelligence gathering, sentiment analysis, etc. Translating external content in order to understand the world creates a big opportunity for MT.
Having a weakness for futuristic technology sessions, I must confess my personal conference favorites were the W1 Multimodal Post-Editing tools session, the C1 American Sign Language session, and all the simultaneous Speech2Text /Text2Speech apps. It is so exciting to see the ideas and the progress, and to imagine what it will be like in a few years when these tools are commonplace!
I hope you’ve enjoyed this AMTA recap as much as I enjoyed attending. Now that I have seen with my own eyes the content and the level of participation, I highly recommend this event, and am really looking forward to future editions. Bravo Team AMTA!
To learn more about AMTA and access the recordings, including SDL’s session (C18),
please click here.
