Community

TAG: A Better Way to Harmonize Terminology and LLMs

Supported by LocWorld

T

erminology is an essential resource for translators — one that helps them maintain a consistent, on-brand voice. The problem is it often doesn’t harmonize with external artificial intelligence (AI) systems — or, for that matter, a business’s other departments.

If you ask Klaus Fleischmann, founder and chief executive officer of Kaleidoscope, “harmonization” is exactly the right way to look at it. He believes terminology systems have more than a few points in common with the music industry’s evolution. After all, you can have the best sound system in the world, but if you’re not connecting it with world-class music, there’s not much of a point, right?

The same holds true for large language model (LLM) translation. It needs the context and guidance of a term base to truly sing. But coordinating that duet requires a little technical innovation.

“There were some attempts [to address this problem],” Fleischmann said. “We all tried prompt engineering, we all tried retrieval-augmented generation (RAG). None of that really worked. We all tried glossaries and machine translation engines. All of that has its problems.”

Fleischmann envisioned a better method, one he calls terminology-augmented generation (TAG). And at LocWorld’s 19th Process Innovation Challenge, Fleischmann demonstrated the seamless process, landing the second runner-up award in the process.

Similar to the smartphone’s consolidation of the purchasing, organization, and playback of a music library onto a single system, TAG’s smart, automated system integrates external data sources and streams terminology to the entire team.

“The marketing person is trying to figure out, ‘How can I stay on brand linguistically?’ The data people are saying, ‘How can I generate correct data?’” Fleischmann said in his presentation. “And of course, we as linguists and terminologists are scratching our heads saying, ‘Why aren’t they looking at us? We have the solution for this — it’s called terminology!”

TAG distinguishes itself from previous attempts to integrate terminology systems by solving a few key pain points. For one thing, it’s a highly structured data model, making all information easy to navigate, locate, and filter. It also takes a flexible approach to formats, ensuring broad compatibility. And it all happens in real time, eliminating the need for uploading.

But these features don’t address the most important issue: feeding that terminology into the LLM. Fleischmann solved this problem with the Model Context Protocol (MCP). Once again, the music industry provides a useful comparison with its pivot to streaming, since MCP works much the same way.
“It basically allows you to connect external data sources to the LLM process,” Fleischmann said. “And one of those is terminology.”

Fleischmann illustrated his point with a simple example of Spanish translation using vela, a word with two distinct meanings. TAG easily retrieves not only the translation but also the definitions, identifying the potential for translation error.

“That’s enough for the LLM or for modern machine translation engines to actually translate it correctly,” Fleischmann said.

Even better, TAG’s direct integration with an enterprise LLM means that the entire workforce is accessing correct terminology. We all understand how difficult it is to add additional steps or tools to long-memorized methods and processes. But because the LLM is the agent accessing the term base, and everyone is interacting with the same LLM, professionals of all stripes will hone their linguistic precision without ever knowing it.

“One of the big problems of terminology is it’s really difficult to promote it in the company and say, ‘Use the term base! There’s so much information in there!’” Fleischmann said. “But by just plugging it straight into the LLM tools, they’re actually using it without even noticing.”

“And of course, it increases the value of the localization team because you provide that data,” he added.

Demonstrating provable value is nothing to sneeze at. After all, words still matter — the trick is communicating that value in a way that resonates. And Fleischmann believes TAG translates that value in the most persuasive terms possible: cold hard dollars and cents.

“It’s a game-changer for the world because it unlocks the savings, and for the localization industry because we stay relevant,” he said.

Related Articles