Google Translate Makes Largest Update in History!
Google has announced a major expansion of its translation capabilities, introducing support for 110 new languages. This brings the total number of supported languages from 133 to 243, fulfilling longstanding requests such as Cantonese, marking the most significant update since the launch of Google Translate.
According to Google, this update relies on the PaLM 2 AI language model. Isaac Caswell, a Google research scholar, highlighted PaLM 2’s exceptional performance in learning highly correlated languages. Examples include Awadhi and Marwadi related to Hindi, as well as Creole languages like Seychellois Creole and Morisien, which are closer to the French language family.
Google now understands Cantonese too. Caswell noted that Cantonese has long been one of the most requested languages for Google Translate. However, due to overlapping with Mandarin in writing, finding appropriate data and training models presented a significant challenge.
In addition to Cantonese, Google Translate now supports other new languages. Caswell mentioned that approximately a quarter of the new additions are from Africa, underscoring Google’s commitment to digitizing African languages.
Caswell revealed that most newly added languages have at least one million users, with some languages reaching hundreds of millions. These additions expand Google Translate’s utility and practicality in multilingual environments.
What is PaLM 2, and why is it powerful? Google explained that the technological backbone behind this language expansion comes from the robust learning capabilities of the PaLM 2 AI language model. This model not only effectively learns and understands new languages but also establishes connections between related languages, thereby enhancing translation accuracy and naturalness.
PaLM 2, released by Google in 2023 as the second generation of large-scale language training models (LLMs), has significantly improved capabilities in areas such as mathematics, logical reasoning, and encoding compared to its predecessor, PaLM 1.
Google’s PaLM 2 is available in four versions—Unicorn, Bison, Otter, and Gecko—tailored for different types of devices. The lightweight Gecko version can even be used offline directly on mobile devices.
PaLM 2 was trained on over 100 languages, specializing in tasks such as natural language understanding, translation, encoding, question answering, summarization, creative writing, mathematical logic, and common-sense reasoning. Particularly in semantic understanding, PaLM 2 excels in interpreting non-literal texts such as riddles and idioms.
Source: The Verge, Google