Webinar: Can we Trust LLMs to do our QE?
5 March 2024
5 - 6 pm CEST

In this TAUS webinar, experts in the field of MTQE will engage in a comprehensive discussion on the merits and drawbacks of two distinct approaches: the open (LLM) approach and the custom-built implementation. Uncover the nuances of LLMs, acknowledging their strengths while addressing their potential pitfalls, particularly in highly domain-specific contexts. Explore the resource and cost implications of fine-tuning, and maintaining LLMs, contrasting this with the more controlled and customizable nature of a custom-built MTQE solution. Gain insights into strategically leveraging LLMs in combination with MTQE to enhance content rewriting and error rectification. Join us as we navigate the landscape of MTQE and make informed decisions for the future of your machine translation initiatives.

2024 will be the year of MT Quality Estimation. Everyone who is using MT - and who isn’t these days? - will now want to use MTQE. Why? Because after all the excitement and hype around AI and LLMs in this past year, we want to put our feet on the ground and build trust in our MT setup. Besides, use cases tell us that one can easily save 30% or more on post-editing costs and efforts. TAUS is already processing around half-a-billion characters each month through our Estimate API, blocking the really bad machine translations from reaching our customers’ customers and flagging the good translations so that they don’t need to be sent to post-editors.

The question though is: what is the best way of implementing MTQE? There is a lot of talk about LLMs being capable now of also predicting the quality of MT. But how reliable, consistent and accurate are the LLMs really? And what about the cost and privacy issues around the set-up of QE with LLMs? The alternative is to implement your own controlled environment with pre-trained custom models, the approach offered by TAUS through the Estimate API.

What we talk about

What is quality estimation and how does it work?

Are large language models capable of doing quality estimation?

What are the pros and cons of using large language models?

Q&A with the expert panel

Konstantin Dranch
Owner | CustomMT
Christian Federmann
Principal Research Manager | Microsoft
João Graça
CTO | Unbabel
Amir Kamran
Solution Architect | TAUS
James Lin
Technical Program Manager | Uber
Jaap van der Meer
Anne-Maj van der Meer
Head of Sales & Marketing | TAUS