In 2005 I published a manifesto under the title Translation-out-of-the-Wall. Translation-as-a-Utilility like the internet or electricity, ubiquitously available to every citizen on the planet, nothing less, was the founding vision and mission of TAUS. The late Hans Fenstermacher, co-founder of GALA, warned me: “Be careful what you wish for, Jaap!” I laughed, but recently I am thinking more often of his foresight. Technology has evolved so fast: more than ‘just’ translation we are now getting addicted to Intelligence-as-a-Utility: IaaU. Browsing through the hundreds of LinkedIn posts and blog articles every week, it is clear that the translation industry has lost its foundation and everyone is struggling to define a new role for themselves.
Over the past decades the job of a localization manager was simple and straightforward: find the TMS of your liking and ensure a successful implementation and adoption throughout your company. And then, once every five years or so, something happened that triggered the search for another TMS: new features, new pricing, new integrations. Now, it is not so simple anymore. The AI revolution has turned everything upside down. Look at the titles of the sessions at the TAUS conferences this year, and you sense the uncertainty:
The Globalization Dilemma: a division between pragmatists and radicals
The LSP Transformation: what is replacing translation?
From TMS to AI Orchestration: Can the TMS infrastructure hold everything together?
Everyone is desperate to define a new destiny. In the pursuit of a new future many localization managers follow the old pattern and open the search for a new TMS, again! Hoping that that will kind of magically resolve all their issues. But it won’t.
The AI revolution, mixed with the geopolitical shifts, creates a much more complex reality than we could imagine until perhaps even six months ago. Everyone, every government, every enterprise, every LSP, is now facing the challenge to take back control of their cultural, linguistic heritage and destiny. The new world order is making it painfully clear that ownership of our language, data sovereignty in a more trendy term, is the real key to our future. In the pre-AI world, companies would hire teams of linguists and translators to preserve their company brands, culture and communications. And some still do. But for most companies it feels like an anachronism to hire translators and linguists, because now there are LLMs that take care of that job. But they can’t.
Large Language Models don’t speak for you. They don’t know your terminology, your style. They have the tendency to be biased in many ways. To the English language as the source of everything. To the western hemisphere as the geography of choice. To axioms voiced by the most active and populous users of the internet. To the politics and preferences of the owners of the LLMs. In their attempt to circumvent these biases, TMS’s and LSPs tap into multiple LLMs hoping to find the right text, translation, quality score or at least the best average. This use of Large Language Models is sub-optimal in a localization environment. Let’s face it: to get machines to really speak our language we will have to feed the algorithms with all our linguistic, stylistic preferences and our own biases.
The term ‘cultural intelligence layer’ started to pop up in LinkedIn posts and blog articles. That sounds fancy and interesting. Indeed, if LSPs and internal localization departments could ensure that AI is not mass-producing content of bad quality by supplying this ‘cultural intelligence layer’, that would be fantastic. In the tumult of the desperate search for the future, many promises are made, halve promises, false promises. Companies jump ahead of themselves with new features and claims. The reality is that we as an industry have not solved the problem yet of ensuring that the AI models understand and speak our companies’ and customers’ language. As long as we haven’t automated, or even half-automated, this cultural intelligence layer, it remains just a concept.
The good news is that there is a way to automate the cultural intelligence layer. There is a path to personalize and purify the translation bots, to make them use the metaphors of fashionista’s, the diplomacy of politicians, the jargon of gamers, to make them use our private company’s terminology and style of communications. It is hard work. I know that, because at TAUS we got our hands dirty: from talking the talk to walking the walk. A few years ago we started putting our massive TAUS data repository to work to build a private model, first for Uber, then for other enterprise customers and LSPs. I call it an art because it takes a lot of practice to tune models to make them speak your language, so to say. And I call it a science, because the NLP team we have assembled to develop the EPIC models all have master and PhD degrees in linguistics and computer science.
What we built is a new category of technology that is now known as Quality Estimation (QE) technology. Gartner Research in a recent report states that this is an essential technology that bridges the gap between MT engines or LLM’s and the translation quality requirements and specific use cases of each individual company and government. Gartner predicts that by 2027 every translation operator in the world will have implemented this new technology.
Quality Estimation technology in combination with Automatic Post-Editing enables localization teams to put their own data to work - translation memories, glossaries, style sheets - and implement their company’s private AI quality companions, allowing them to scale up production very fast while maintaining their unique value proposition of quality.
And so, we are getting closer, step-by-step, to what I wished for in 2005 at the foundation of TAUS.