Online gaming platform Roblox has unveiled an innovative AI-driven "unified translation model," enabling seamless real-time text conversations among players speaking different languages.
According to Roblox Chief Technology Officer Dan Sturman, the development involved creating an in-house large language model (LLM) to achieve a base latency of 100 milliseconds, ensuring instant communication for users.
Sturman highlighted the scenario of users chatting in different languages without realizing it, emphasizing the capability of the LLM to translate text-based messages in real time across 16 supported languages.
Roblox's journey in building the "unified translation model" involved overcoming challenges in designing a system for independent translation among all 16 languages and ensuring swift real-time support for chats. The solution involved creating a unified, transformer-based translation LLM to handle all language pairs within a single model, as opposed to building separate models for each language pair.
The process began with Roblox developing a transformer-based LLM, trained on a mix of public and private data. Expert translation apps were then employed to train the model for each language. Addressing challenges in less common translation pairs, Roblox implemented "back translation" and utilized a quality estimation model to assess translation understandability.
The translation model was further refined to comprehend human slang, with human evaluators continuously updating the system with current popular terms in multiple languages.
In final testing, Roblox observed enhanced engagement and session quality, indicating the positive impact of the new translation system on its platform. With 70 million daily active users from over 180 countries exchanging over 2.4 billion messages daily, Roblox continues to explore advancements, with CEO David Baszucki expressing aspirations for metaverse interoperability and the seamless transfer of digital assets across independent platforms.