Roblox Developing a Multilingual Translation Model for Real-Time Chat

Roblox Developing a Multilingual Translation Model for Real-Time Chat

Roblox, a gaming company, is creating a multilingual translation model with artificial intelligence (AI) to facilitate real-time chat interactions on the platform.

The company said in a blog post that it used Roblox’s large language model (LLM) with sophisticated functionalities to accommodate 16 languages to deliver the feature. Thanks to its AI-based translation tool, Roblox’s 70 million daily users can converse with one another in real time without any lag.

Multilingual translation

Roblox’s LLM reportedly achieved the milestone because of its processing speed, with a latency of under 100 milliseconds. In addition to its processing speed, the AI feature supports “Roblox-specific language, colloquial language, and abbreviations.

Roblox’s development has multiple advantages to creating individual LLMs for each language pair since it houses all the language pairs in a single interface. The unified, transformer-based translation offers time and cost-saving benefits as it maintains translation accuracy.

According to Daniel Sturman, the Chief Technology Officer at Roblox, this architecture allows for better utilization of resources since each expert has a different specialty, which leads to more efficient training and inference without sacrificing translation quality.

However, the in-house-developed LLM supports 16 languages, including English, Japanese, French, Polish, Vietnamese, Thai, and more, quickly translating between any pair of languages. Users can still view the original message by clicking the symbol next to the player’s name, while banned and cursed words will still be blocked.

Roblox’s model

Upon closer inspection, Roblox’s model shows that the unified system can lean on the similarities between multiple languages to increase accuracy. Preliminary research indicates that the model can accurately identify input in various languages and provide a reasonably comprehensive translation.

Roblox’s technical team used several training data sets to ensure that the translation model was well developed. The team started with open-source data and added chat translation results and in-house data while depending on back translation to handle complicated translation pairs.

The Web3-based gaming company said that to ensure a safe gaming experience for users, the LLM will automatically block phrases violating its policies.

According to Sturman, safety and civility are at the forefront of everything they do at Roblox, so this was an essential piece of the puzzle.

Roblox’s development plans

Roblox’s CTO disclosed that the firm will steadily gather user feedback on the platform to improve its AI-based translation feature. While considering adding new languages, the technical team revealed that regular updates will be made to keep up with new slang.

According to Sturman, the hope is for the translator model to move past just translating text chats eventually. He added that in the future, they could use AI to translate non-compliant words to compliant words or throw it at voice chats for real-time voice translation.

Other companies are also developing AI translation models. SeamlessM4T, a speech-to-text and text-to-text translator that handles nearly 100 languages, was released by Meta.

In addition to translating into about 100 languages, Google’s Universal Speech Model has been used to translate captions on YouTube.

Image credits: Shutterstock, CC images, Midjourney, Unsplash.