The online gaming platform Roblox has developed a new "unified translation model" based on artificial intelligence, allowing players to communicate with each other through text in real-time, despite speaking different languages. According to Roblox's Chief Technology Officer Dan Sturman, the new system required the company to create its own LLM (large language model) to be able to translate text messages with a basic delay of 100 milliseconds, making conversations for users instantaneous.
The LLM translates text messages in real-time across 16 supported languages, enabling players from different countries to freely communicate with each other. To achieve this, Roblox created a unified LLM for translation based on the transformer, which processes all language pairs in a single model, avoiding the need to create a separate model for each language pair.
The model was trained on public and private data and further trained on each individual language using "expert" translation applications. Roblox also used the "back translation" method to improve the quality of translation for less common language pairs and engaged specialist evaluators to translate popular and trending terms for each language.
As a result, the new translation system has facilitated more active interaction and session quality for users of the Roblox platform, which boasts 70 million active users from more than 180 countries worldwide, exchanging more than 2.4 billion messages daily.
