EasyTranslate thinks augmenting LLMs with humans will give it an edge over pure AI translation services

You might think new generative AI startups like Eleven Labs are the hottest game in town for translation services. But voice translation was long ago preceded by another market, targeted some time ago by startups: content translation. Any company with an international presence needs to have their content translated around the world, so this remains a big market. This was evidenced by the $106 million raised to date by the likes of Unbabel in Portugal (which last raised $60 million). 

EasyTranslate, which specializes in content translation, has been around since 2010, using machine learning models to identify which freelance translators were best suited to translate specific types of content. But now it’s headed in a new direction with a new, generative AI-driven platform that it calls ‘HumanAI’.

“We have pivoted the whole business model from a human service-based business model towards being an AI technology provider, driving down the cost and speeding up the process,” the company’s founder Frederik R. Pedersen told TechCrunch. 

Most translation services offer machine-translated content, with a small portion edited by humans. However, translators often must assess the entire machine-generated translation to understand the context and make sense of the content. EasyTranslate’s HumanAI platform flips this on its head, absorbing content, blending it with large language models (LLMs) and employing short-term memory in the LLM to translate content more accurately. What’s more, it will only involve humans where it needs to, thus reducing translation times and costs.

To do this, HumanAI uses a mix of LLMs, including the one offered by OpenAI, as well as its own recommendation systems. The platform runs off its own algorithms and customer data to provide customized content translation.

The secret to the pivot, Pedersen said, is using LLMs to generate short-term memory so the platform can read a translation in generic English and turn it into specific English. It “vectors” content into a database, enabling it to do a semantic search and find similarities between content, which is then used to create a short-term memory with an LLM (this is also referred to as retrieval augmented generation).

This means the platform can use any number of LLMs to translate between, for example, the English used in marketing copy or English employed in finance reports, and preserve the meaning in the text all the while. 

“We can combine the more traditional, neural machine translation engines with customer-specific data to create a foundation for the localization and translation process. So, moving from generic language towards customer-specific language, for instance,” he said.

Why is that important? Pedersen explained: “You might get a grammatically perfect machine-based translation, but it still may not sound right. So we identify which part of the content has a low confidence score and then use humans to correct it. The combination massively increases our productivity.”

Pederson claimed HumanAI can drive-down translation costs by 90%, and ends up pricing its services at €0.01 per translated word. Its customers include global businesses such as Wix and Monday.com. 

And pricing is an especially crucial puzzle to solve in this space because companies have a great deal of content that needs translating.

“If you look at Adobe, they have a full team just looking at how the terminologies align across markets. And if we look at global brands, there’s a significant amount of effort put into making sure that you are perceived in the right way locally,” Pedersen said.

The question is, though, what will help EasyTranslate compete against pure-play AI-based solutions, which are likely to get better with time?

“Our goal is not to become a pure AI [service]. I think our goal is to create the added value of having humans combined with AI, and provide this service to customers. AI still needs human feedback to be improved,” he said.

“It’s one thing to say you would like to implement all content creation, all translation, and another to make sure that you can actually control the model. You have to have some humans to control the models, because humans are not machines and language changes constantly.”

EasyTranslate has raised a total of €3 million to date, and is backed by private equity, debt financing, some angel investors in Copenhagen and the Danish Innovation Fund.