Roblox AI introduces a unified translation model allowing real-time text communication among players speaking different languages.
Powered by artificial intelligence translation model, This has made it possible for players to communicate with one another through text in real time, even if they speak different languages.
In a statement released on February 5, Roblox chief technology officer Dan Sturman explained that the company constructed its own in-house LLM to translate text-based messages at a base latency of 100 milliseconds. Roblox constructed its own in-house LLM to give users the impression that they are having real-time conversations.
“Imagine discovering that your new Roblox friend, a person you’ve been chatting and joking with in 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.
“This is happening without either of you even realizing it.” Despite the fact that it has already automatically translates its in-experience information, Sturman mentioned that the company wished to “go beyond translating static content into experiences.”
According to Sturman the two most significant challenges in developing the translator were the design of a system that enabled translation between all sixteen languages independently and the development of a system that was fast enough to support real-time chats. This required a novel approach to the construction of the translator’s own large language model (LLM).
“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.”
The development of Roblox’s artificial intelligence translator began with training a transformer-based LLM on both public and private data. Subsequently, Roblox transferred the LLM to a variety of “expert” translation applications, which were responsible for training the model in each specific language.
According to Sturman, “less common” translation combinations, such as French to Thai, were difficult to translate because there was a dearth of high-quality data. As a result, Roblox had to use “back translation,” a process where they translated messages into the original language and verified their accuracy with the source text.
The quality estimation model analyzes the translations, giving priority to the target audience’s understanding. Roblox trained the model to comprehend human slang and brought in human assessors to interpret “popular and trending terms” for each language.
Furthermore, the model underwent training to comprehend human spoken language. Sturman further stated that human reviewers continuously upgraded the system to ensure it remained current with the most recent words.
Roblox discovered, during the last stage of testing, that the improved translation system resulted in “stronger engagement and session quality” for users of its site.
Currently, Roblox boasts over 70 million daily active users spanning more than 180 countries worldwide, with over 2.4 billion messages sent and received each day.
During the month of November in the previous year, David Baszucki, the CEO of Roblox, expressed his desire for interoperability and stated that he believes all users of the metaverse should have the capacity to transfer nonfungible tokens (NFTs) and other digital assets between several independent platforms.
