You might imagine that latest generative AI startups like Eleven laboratories are the most popular marketplace for translation services. But language translation was long preceded by one other market that startups began specializing in a while ago: content translation. Any company with a world presence must have its content translated worldwide, so this stays a giant market. This has been demonstrated by the $106 million raised to this point by corporations like Unbabel in Portugal (which recently raised $60 million).
EasyTranslatewhich focuses on content translation, has been around since 2010. It uses machine learning models to work out which freelance translators are best suited to translate certain varieties of content. But now it's entering into a unique direction with a brand new, generative AI-driven platform called “HumanAI.”
“We modified all the business model from a human services-based business model to an AI technology provider to scale back costs and speed up the method,” company founder Frederik R. Pedersen told TechCrunch.
Most translation services offer machine-translated content, with a small portion edited by humans. However, translators often need to review all the machine-generated translation to grasp the context and make sense of the content. EasyTranslate's HumanAI platform turns this on its head: it takes content, combines it with large language models (LLMs), and leverages the short-term memory within the LLM to translate content more accurately. What's more, humans are only involved where obligatory, reducing translation time and costs.
To do that, HumanAI uses a combination of LLMs, including the one offered by OpenAI, in addition to its own suggestion systems. The platform uses its own algorithms and customer data to supply tailored content translations.
The secret of the pivot, says Pedersen, lies in using LLMs to create short-term memory, in order that the platform can read a translation usually English and convert it into specific English. It “vectorizes” content right into a database, which allows it to perform a semantic search and find similarities between content, that are then used to create short-term memory with an LLM (this can be generally known as Retrieval prolonged generation).
This implies that the platform can use any variety of LLMs to translate, for instance, between English in marketing texts and English in financial reports, while all the time preserving the meaning of the text.
“We can mix the more traditional neural machine translation engines with customer-specific data to create a foundation for the localization and translation process. For example, we will move from generic language to customer-specific language,” he said.
Why is that this necessary? Pedersen explains: “You can get a grammatically correct machine translation, but it surely still doesn't sound correct. So we discover which a part of the content has a low confidence rating after which have humans correct it. This combination increases our productivity enormously.”
Pederson claimed that HumanAI can reduce translation costs by 90% and calculates that its services cost €0.01 per word translated. Its clients include global corporations corresponding to Wix and Monday.com.
And pricing is a very difficult puzzle on this area because corporations have a considerable amount of content that should be translated.
“At Adobe, there's a complete team dedicated to how terminologies align across markets. And after we have a look at global brands, there's lots of effort put into ensuring they're perceived accurately locally,” Pedersen said.
The query, nonetheless, is how can EasyTranslate compete against pure AI-based solutions, that are prone to get well over time?
“Our goal isn’t to develop into a pure AI service provider. I believe our goal is to create the added value that humans mix with AI and offer that service to customers. AI still needs human feedback to be improved,” he said.
“It's one thing to say you need to do all of the content creation and translation yourself, but it surely's one other to ensure that you possibly can actually control the model. You must have some people controlling the models because people will not be machines and language is always changing.”
EasyTranslate has raised a complete of €3 million to this point and is supported by private equity, debt, some angel investors in Copenhagen and the Danish Innovation Fund.
