Lessons from LanguageWire’s Machine Translation Journey

Development of solutions and technology for machine translation (MT) is becoming a differentiating feature across the language industry. Case in point: Danish language service provider (LSP) LanguageWire, which translated almost no content with MT in 2018.

By 2021, however, the company used MT to translate more than a billion words — “quite a substantial ramp-up” in the words of Roeland Hofkens​, LanguageWire’s Chief Product and Technology Officer, speaking at SlatorCon Remote December 2021.

That said, the company already had a solid foundation to build on — namely, an automated workflow and a proprietary translation management system (TMS). Now, LanguageWire looks for openings in the workflow where automation can free up an employee’s time for work that cannot be left to machines.

A focus on data security and custom MT led LanguageWire to create its own proprietary engines with the help of a team of scientists and engineers. This more flexible system has the added benefit of working five times faster than the open-source MT provider LanguageWire first used, according to Hofkens​.

The company’s translation productivity (a.k.a. CAT) tool already measures metrics such as translation speed and edit distance, painting a more precise picture of what translators do and how long it takes them. Now, the company is building a model that draws on data from past jobs to predict MT quality and effectiveness for a new task.

“We’ve done the simulation on that,” Hofkens explained. “If we apply a higher discount on jobs where vendors have to do almost nothing, and then have no discount or a very small discount on places where the MT does not work, that brings a lot more fairness into how we pay our translators.”

Beyond the details of linguists’ work, LanguageWire has a platform to collect, analyze, and visualize data from all along the project pipeline, from product orders to discounts, allowing decision-makers to compare MT and non-MT jobs.

Hofkens said these insights play a critical role in the company’s MT strategy, as he thinks they should at other LSPs. “If you cannot measure everything that’s happening in that end-to-end flow, then you’re running your MT operations completely blindfolded,” he pointed out.

Watch “Human-centric Machine Translation at Scale” with Roeland Hofkens, and the full SlatorCon Remote December 2021 event, on demand here.