[GALA Valencia 2024] How to build and offer new AI-based services – our experiments and results

22 Apr 2024

This event has expired, video available

To view the recording, you must be logged in with a GALA Member account or have purchased the webinar.

We know that depending on how you look at it, ‘generative AI’ is a great tool, resource or assistant in our industry. Is this the actual disruption that even its predecessor, NMT, has promised to be? The jury is still out. What we know for certain is that generative AI is capable of solving problems that were outside of the automation domain until now. ‘Translation’ is no longer the only service we could potentially offer. Where before we as an industry might have had a monopoly on linguistic services, now we see these same linguistic operations being done largely in-house or solely by developers.

This poses the question: how can you make sure you don't fall behind, and offer similar or better linguistic services using generative AI? Everyone has played with the ChatGPT interface. We all know that you can use it to rewrite content for different audiences, that you can ask it to analyze texts or that you can generate ‘completely new’ content. We wanted to go a bit further than this low-hanging fruit. Over the course of last year, we have looked at the capabilities of large language models, have built workflows that had them operate on actual content, have tweaked parameters and even built workflows that span multiple different models with different prompts. We found that the best results are achieved when you actually chain the content through different actions, keep the prompts focused and tweak the additional parameters that these models offer. We also recognize the differences between the models that different companies provide and can elaborate on the pros and cons.

We have built interesting new workflows that we would like to share as they could help inspire everyone to innovate and advance their service offering. The examples we can show include but are not limited to: automated style-matching, document-level QA, automated post-editing, automated LQA, automated human-translation verification, platform-specific localized repurposing and sensitive content obfuscation, among others!

Host organization: Globalization and Localization Association

Event Speakers

Mathijs Sonnemans
Blackbird.io

Solution architect, software engineer and writer with a background in theoretical computer science and philosophy. My ambition is to bring serious innovation to the localization industry.