Gartner predicts, in their recent report Improve Translation Quality at Scale, that by 2027 every translation operator in the world will have Quality Estimation (QE) implemented coupled with their MT and post-editing efforts. That’s understandable: nobody wants to be left out of this revolution in translation quality evaluation. The introduction of QE technology marks a fundamental shift from reactive quality evaluation to proactive quality management. Besides, the potential business benefits are too big to be ignored.

What is a QE model?

What can a QE model do, in particular in relation to MT? A QE model is an AI model that is trained for the specific task of scoring the quality of output from automatic translation. No matter how good LLMs and MT become at generating multilingual content, there will always be a need for quality control.

QE models are trained to detect translation errors. And when QE models are customized for a particular domain or use case, they are the perfect AI Quality Companion for professional service providers. Understanding and maintaining the style and terminology of their customers in all the languages in which they do business has always been the key differentiating factor for Language Service Providers.

With QE they can scale up their capacity significantly, while maintaining or actually improving their quality control for their customers. 

But how to get started with QE? Build, Buy or Partner?

Building your own QE model. It sounds simple perhaps: just use an LLM! Assemble all of the post-edits of the last couple of years and feed these into a model. At first sight the results may look good. But how do you ensure that the scores are consistent and dependable across all the different languages that you need to manage? What about privacy and data protection? You will find out quickly that you will need dedicated NLP engineers in-house to get the QE models to work better. (See this related blog on the challenges of using LLMs for QE). And still, it is taking too long for the custom models to be finished and ready to go into production. The cost of building your own QE solution will continue to increase. Unless you have very deep pockets and it’s crucial for your company to own your technology, you are better off buying QE.

Buying QE. As Gartner states in their report, TQE technology is just emerging and there is not an abundance of vendors to choose from. When monitoring QE solutions, Gartner recommends to ensure independence from specific Translation Management Systems and MT providers. This gives users of QE more control over their translation workflows, levels of automation and required quality standards for each use case. It is no surprise that Gartner names TAUS as independent QE provider of choice (out of only two!) in their market report. TAUS started developing its QE solution already in 2021, working closely with Uber as a launching customer. In the past four years, TAUS has built a reliable and scalable QE platform with multiple global enterprises and dozens of LSPs sending millions of words through the EPIC API every day for real-time quality scoring.

Partnering with TAUS. Not just Gartner, but also CSA Research analysts Arle Lommel and Alison Toon flag independence as the number one decision criterium when selecting a QE solution in their TAUS EPIC Vendor Briefing report. CSA highlights the ‘white-labeling’ option (offering the QE feature under your own brand) offered by TAUS as a compelling unique feature for Language Service Providers. TAUS has now further developed this feature into a Partner Program, leveraging the EPIC API and skills of the TAUS NLP team with the unique knowledge that LSPs have of their customer’s domain-specific language. Over many years, LSPs have built up a strong quality reputation with their thorough understanding of the vocabularies, style and jargon that are common in their customer’s industries or locales. Through this partnership we try to transform this knowledge as much as possible into a Specialized QE model, allowing the LSP to scale their business, solidify their reputation in the age of AI and generate revenues from the QE technology as well.

What qualifies for the EPIC Partner Program

The generic QE model that TAUS offers out-of-the-box covers more than a hundred languages. This generic QE model already works quite well for many applications and use cases. For specific customer use cases TAUS recommends the training of custom QE models. The EPIC Partner Program focuses on making Specialized QE models available for specific domains, covering the area in between the generic QE model and the custom QE model (that is tailor-made per customer). Specialized QE models can be made available for a wide range of domains. Think for instance of radiology, nuclear energy, pharmacy, games, legal, etceteras. Or we may train a Specialized QE model for a specific language combination. See for example our announcement of TRSB and TAUS joining forces to launch an AI Quality Companion for the Canadian market.

If you want to know more about the TAUS EPIC Partner Program for LSPs, please get in touch with us.