Microsoft is employing a new family of AI models known as Z-code, which provides the quality and performance of large-scale language models while also being more efficient.
We are making incredible progress with Z-code since we are leveraging both transfer learning and multitask learning from monolingual and multilingual data to develop a state-of-the-art language model that we presume has the best mix of quality, performance, and capability that we can offer to our customers, said Xuedong Huang, Microsoft technical fellow, and Azure AI chief technology officer.
Part of what distinguishes Z-code is its ability to derive solutions from model portions rather than the entire thing, making Z-code usage scalable and efficient. It’s being used within Azure AI via the Microsoft Translator app, where it’s helping to expand the Translator’s capabilities for less common languages while also enhancing the results of even the most common translation requests for popular languages like English and French.
For specific numbers on how the model family is assisting Microsoft, the company provided the following hard figures regarding Z-effects code’s on translation: The models enhanced English to French translations by 3.2 percent, English to Turkish translations by 5.8 percent, Japanese to English translations by 7.6 percent, English to Arabic translations by 9.3 percent, and English to Slovenian translations by 15%.
In other words, if we want to translate something to or from English, chances are Z-code will assist us in that process.
