Microsoft improves AI translation with Z-Code

Microsoft has released an update to its translation service thanks to new machine learning technologies. Improved translation between many language pairs. TechCrunch reports: Based on a project Z-code that uses a “pre-expert blend” approach, this new model scores 3% to 15% better than the company’s previous model during blind evaluation. Z-Code is part of Microsoft’s broader XYZ-Code initiative, which is examining combining models for text, vision, and audio across multiple languages ​​to create more powerful and useful AI systems.

“Mixture of Experts” isn’t an entirely new skill, but it’s especially useful in the context of translation. At its core, the system essentially divides a task into multiple subtasks and then delegates them to smaller, more specialized models called “experts”. The model then decides which tasks to delegate to which experts based on its own predictions. You can think of it as a model that has been greatly simplified and contains several more specialized models.

Microsoft improves AI translation with Z-Code

Source link Microsoft improves AI translation with Z-Code

Back to top button