Roblox Develops a Unified, Transformer-Based Translation Language Model (LLM) for its AI Chat Translation Feature.
Roblox has built an artificial intelligence (AI) tool that can translate supporting languages in chats in almost real-time. There is a latency of roughly about 100 milliseconds, claims the company. The online gaming platform, which has more than 70 million daily active users, created a native large language model (LLM) to build the tool that can translate any combination of 16 languages. The foundational model behind the feature puts a heavy emphasis on accuracy, fluency, as well as the context of the rest of the sentence to find the best translation.
Announcing the new feature in a post, Roblox Chief Technology Officer Daniel Sturman said, “Using AI to automate real-time translations in text chat removes language barriers and brings more people together, no matter where they live in the world.” The tool currently supports English, Chinese (Simplified and Traditional), French, German, Indonesian, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai, Turkish, Polish, and Vietnamese.
The feature activates on its own in chat boxes. When multiple users type in different languages, the AI will automatically translate them in real time to the user’s default language. A translation icon is placed in front of the sentence, clicking which turns the language back to the original.
During the development of the AI tool, the online gaming platform encountered a unique challenge: while translating between two languages was straightforward, handling a combination of 16 languages proved to be more complex. Instead of creating 256 separate models (16 x 16), the company adopted an innovative approach. Sturman explained that they employed a unified, transformer-based architecture for the Language Learning Model (LLM), where all translations are governed by the source sentence and the target language. “It’s akin to having multiple translation apps, each specializing in a group of similar languages, all accessible through a single interface,” he added.
This system enables several distinctive features. Besides maintaining a strong focus on accuracy, fluency, and context, the AI tool can translate sentences containing a mixture of different languages. Moreover, human evaluators were enlisted to instruct the model on contemporary slang and trending phrases in each of the 16 languages. This ongoing process ensures that the LLM remains up-to-date.
Furthermore, the company employed a ‘back translation’ technique for less common translation pairs, such as French to Thai. In this approach, Roblox first translated the content and then translated it back to the original language. The source material was then compared to the back-translated version to identify inaccuracies. Labelled data was subsequently utilized to augment the volume of translation data for the model’s learning. Sturman also revealed that Roblox is exploring automatic voice chat translations for its users.