Online gaming platform Roblox has developed a new AI model that allows players to chat with each other in real-time, even if they speak different languages. Currently available in 16 languages, the AI-powered translator enables users to type in their native language and easily converse with others.
Roblox’s Chief Technology Officer, Dan Sturman, stated on February 5 that the company developed its own in-house large language model (LLM). This model allows for text-based message translation with a base latency of 100 milliseconds, providing users with instant conversations.
“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,” said Sturman.
While Roblox already has automatic translation capabilities for its in-experience content, the goal was to expand translation to live conversations.
Stuntman said that the two main challenges in developing the translator were creating an independent translation system that worked for all 16 languages and designing a fast enough system for real-time chats, which required a novel LLM architecture.
“To achieve this, we could have built out a unique model for each language pair (i.e., Japanese and Spanish), but that would have required 16×16, or 256 different models. Instead, we built a unified, transformer-based translation LLM to handle all language pairs in a single model.”
Roblox’s AI translator started with the development of a transformer-based LLM that was trained on both public and proprietary data.
Following development, Roblox entrusted the large language model (LLM) to a blend of “expert” translation applications. These apps trained the model in various languages individually.
Sturman pointed out that translating “less common” language pairs, like French to Thai, posed difficulties due to insufficient high-quality data. This led Roblox to employ “back translation,” a method involving translating messages back into the original language and comparing them with the source text to ensure accuracy.
The translations are then run through a quality estimation model, which prioritizes the readability and comprehension of the translated text.
Additionally, Roblox also had human evaluators help train the model by translating popular, trendy phrases into each language. So the model can understand idiomatic or slang language.
Sturman emphasized that this is an ongoing process, requiring human evaluators to constantly update the system to keep up with the latest trends.
Final-stage testing showed that the new translation system led to more engagement and higher-quality sessions for Roblox’s users.
Roblox has 70 million daily active users from over 180 countries, exchanging more than 2.4 billion messages every day.
David Baszucki, the CEO of Roblox, expressed his vision for interoperability in November 2023, advocating for the seamless transfer of nonfungible tokens and other digital assets across various independent platforms within the metaverse.
Read also: Report: Q3 blockchain gaming Unique Active Wallet surpasses Q2
