Author: Konstantin Dranch
1. Bigger, Bolder Programs
Salvo Giammarresi (AirBnB) and Rachel Carruthers (Canva) paved the way in 2021 with well-publicized programs counting 60+ and 100+ languages. Particularly, the news of Airbnb’s Translation engine made the company’s stock price jump 21% to the highest valuation in history. With daring examples in the public eye, other localization leaders will have an easier time persuading stakeholders to go for plans of scale that capture the imagination of users and investors alike.
- Automated localization as user experience & global presence enabler
- MT increases scaleup capitalization
2. Take Back the Benefit
Very often in 2021 enterprise buyers who put effort into machine translation and trained a fleet of models struggled with converting tech improvements to financial gains due to the resistance of their supplier LSPs.
While translation companies brought translator rates down 50-75% with post-editing, buyers hardly received any benefit. We saw this effect of organizational readiness lagging behind tech readiness with multiple top brands.
This is about to change. Like with translation memory 20 years ago, savvy buyers will fragment their supply chains, and move from single- to multi-vendor relationships to take ownership of the benefits brought by technology. This will require defining quality levels and institutionalizing post-editing quality programs. Hopefully, in this move buyers can be mindful of professional translators and improve their livelihood by switching to more direct relationships with them.
3. Crowd-editing Enabling Machine
Website internal pages, tech docs, product catalogs, and forums present opportunities to optimize budgets by moving from manual to machine translation.
To eliminate the risk of embarrassing errors, however, buyers will return to a crowd-sourcing approach. Crowdsourcing has been all the rage 5 years ago, and then it slowed down due to the difficulty of motivating enthusiasts to cover the last mile, typically the most boring translations.
Today, the technology takes care of this:
- crowd workers are available via APIs
- MT goes to the last mile
Compared to crowdsourced translation, Crowd-Editing requires significantly less effort, and so just a few enthusiasts can cover a lot of ground.
4. Language Variants
Mainstream MT is weak in language variants. For example, Google Translate does not support UK English and Latam Spanish. And let us not get started on Catalan and Swiss German.
This means that translators working in these economically important languages have to routinely deal with issues in vocabulary, units of measure, date and time formats. That’s a lot of tedious corrections which eat up time and energy.
Language variants can be fixed by adding a secondary MT engine that translates US to UK and Continental Spanish to Latam. A lot of such optimization work is necessary this year.
5. Stronger Adoption of Open-Source MT
Training your own MT engine has never been easier. Students do that in classes by downloading OpenNMT and MarianNMT from Github. Baseline training data is abundant with terabytes of translations in major languages downloadable from the EU-funded project websites. No longer the monopoly of search engines, the treasure of the Internet is up for grabs.
Moreover, Facebook AI has published free multilingual models that won the influential WMT competition. These models are available for download.
Buyers who want to own their MT and eliminate the dependency on Google and the rest of Big Tech will have an easier time and will require a smaller investment to do so.
Let’s discuss how to improve your strategy together, book a briefing with us!