Most people who use AI translation tools do so for common. Relatively unimportant tasks, such as understanding a single phrase or quote.
These basic services won’t be enough for a business offering technical documents in 15 languages. But Lengoo’s own machine translation models can do. And with a new $20 million B round, they can build a significant lead.
The translation business is big, in the billions, and it’s not going anywhere. It’s just too common a task to need to release a document, software, or live website in multiple languages—perhaps dozens.
These days, this work is done by translation agencies that employ expert speakers to ensure high-quality on-demand translations. The rise of machine translation as an everyday tool hasn’t affected them as much as you might think.
Because the casual Portuguese user using Google’s built-in website translation on a Korean website is a very niche case. And things like translating social media posts or individual sentences aren’t really something that you could or would like to hand over to the professionals.
In these familiar cases, “good enough” is the rule, because the bare meaning is all anyone really wants or needs. But if you’re releasing a product in 10 different markets where 10 different languages are spoken. It won’t be enough for the instructions, warnings, legal agreements, or technical documentation to be perfect in one language and just okay in the other nine.
Lengoo started from a team working to automate this workflow between companies and translators.
“The next step to take was to automate the translation itself,” said CEO and founder Christopher Kränzler. “We’re still going to need people in the loop for a long time — the goal is to get the models to a level where they’re actually usable. And there’s less translation by a human.”
As machine learning capabilities continue to improve, this is not an unrealistic goal at all. Other companies have gone down this path – such as DeepL and Lilt, which proved their cases by demonstrating major improvements over the Google and Microsoft frameworks. But never claimed to remove humans from the process.
Lengoo repeats its work by focusing on speed and specificity – that is, it creates a language model that integrates all the jargon. Stylistic preferences and formatting requirements of a given client. To do this, they build their own language model by training it not only with the customer’s own documents and websites, but constantly adding feedback from the translation process itself.
“We have an automated training pipeline for the models,” Kränzler said. The more people who contribute to the correction process, the faster the process will be. In the end, we will be about three times faster than Google or DeepL.”
A new client can start with a model fit on a few thousand documents from the last few years. But whenever the model produces text that needs to be corrected, it remembers that particular correction and integrates it with the rest of its training.
While it can be difficult to objectively quantify the “quality” of a translation, it’s not a problem in this case because working as a human translator’s tool means that quality control is built right into it. How good a translation is can be measured by the “correction distance”. ”, basically the amount of changes one has to make to the text suggested by the model. Fewer corrections means not only better translation, but faster translation, which means both quality and speed have objective metrics.
The improvements won over customers who had been concerned about excessive automation in the past.
“There was resistance in the beginning,” Kränzler admitted. “People are turning to Google Translate for their daily translations. And seeing the quality improve – they and DeepL have really educated the market. People now understand that if you do it right, machine translation works for professional use. A large customer can have 30, 40, 50 translators and each of them has their own style. We can say that we are faster and cheaper, but also that the quality, in terms of consistency, is going up.”
While customizing the model using client data is hardly a unique approach, Lengoo seems to have built an edge over competitors. And slower big companies that can’t improve their products fast enough to keep up. And they intend to consolidate this leadership by overhauling their technology line.
The problem is that due to the reliance on more or less traditional machine learning technologies, the crucial translator-AI feedback is limited. How quickly the model updates depends on how much it’s used, but you won’t be retraining a large model just to integrate a few hundred more words of content. Retraining is computationally intensive, so it can only be performed sporadically.
However, Lengoo plans to build its own, more sensitive neural machine translation system that integrates different channels and related processes. The result wouldn’t improve in real time, exactly, but it would contain the latest information much faster and in a less involved way.
“Think of it as segment-by-segment improvement,” said applied research lead Ahmad Taie (segments vary in size, but are generally logical “chunks” of text). “You translate one segment and in the next you already have model improvements made.”
Of course, the key to keeping clients hooked is to make this key product feature better, faster and easier to implement. And while there will likely be strong competition in this space, Kränzler doesn’t expect it to come from Google or any of the existing large companies, which tend to take an acquisition-and-integration approach rather than agile development.
As for human expert translators, the field won’t replace them, but it can eventually increase their effectiveness by an order of magnitude, which can reduce the workforce there. But if international markets continue to grow and so does the need for professional translation, they may just keep up.
The $20 million round, led by Inkef Capital, will allow Lengoo to jump into North American markets as well as other European markets and integrate with more enterprise packages. Existing investors Redalpine, Creathor Ventures, Techstars (from whose program the company was born) and angels Matthias Hilpert and Michael Schmitt joined the round, along with new investors Polipo Ventures and Volker Pyrtek.