Roblox, a popular online gaming platform, has introduced an innovative AI technology named the “unified translation model” to help players communicate across different languages in real-time via text. This advancement was announced on February 5 by Dan Sturman, the Chief Technology Officer at Roblox. The technology is based on a custom-built large language model (LLM) that allows for translating text messages with a minimal delay of 100 milliseconds. This makes conversations appear seamless and instantaneous for users.
Dan Sturman illustrated the impact of this new system with a scenario: “Imagine discovering that your new Roblox friend, a person you’ve been chatting and joking with a new experience, is actually in Korea — and has been typing in Korean the entire time, while you’ve been typing in English, without either of you noticing,” he explained.
How Roblox Developed the “Unified Translation Model”
Roblox aimed to enhance user experience by translating static content within games and enabling real-time text conversations across different languages. Sturman highlighted the main challenges in developing this translator: designing a system capable of translating independently among 16 languages and ensuring the translation process was quick enough to support real-time interaction. This required an innovative approach to creating a singular, transformer-based translation LLM to manage all language combinations within one model rather than developing 256 distinct models for each language pair.
The development process started with Roblox creating a transformer-based LLM using public and proprietary data. Subsequently, this model was refined with the help of various expert translation applications to improve its accuracy for each language.
For less common language pairs, such as French to Thai, Roblox employed a “back translation” method to enhance the quality of translations. This involves translating a message back into the original language and comparing it with the source text to ensure accuracy. Additionally, a quality estimation model was used to prioritize the clarity and understandability of translations.
Roblox involved human evaluators in interpreting slang and current terms for each language to make the translations more relatable and accurate, ensuring the system remains up-to-date with evolving language use.
Sturman mentioned that this process is continuous, with updates regularly integrated to maintain the system’s relevance with the latest phrases and terms.
Impact and Future Aspirations
In its final testing phase, the new translation system significantly enhanced engagement and the quality of user sessions on the platform. With its 70 million daily active users from over 180 countries and more than 2.4 billion daily message exchanges, Roblox sees this as a step towards fostering a more inclusive and connected user community.
In November 2023, Roblox CEO David Baszucki expressed his vision for interoperability within the metaverse, advocating for the ability of users to transfer nonfungible tokens (NFTs) and other digital assets across various platforms, further emphasizing the company’s commitment to innovation and user-centric design.
This development illustrates Roblox’s dedication to enhancing user experience and highlights AI’s potential in bridging communication gaps and fostering global connections.
