TransAgents, a new system, leverages multiple LLMs to achieve superior results.
A groundbreaking new approach to literary translation has been developed by researchers from Monash University, the University of Macau, and Tencent AI Lab. AI system that can translate literary works with a level of sophistication previously unseen in machine translation.
TransAgents, a multi-agent system, employs multiple large language models (LLM) to imitate the roles of human translators, editors, and localization specialists. This collaborative approach allows the AI to capture the nuances, cultural context, and individual style of the original author.
“Prompting a single LLM to translate literature often results in subpar quality,” explained Minghao Wu, one of the lead researchers. “By breaking down the complex task into smaller, more manageable subtasks and assigning them to different agents, we’ve been able to achieve far more impressive results.”
To train TransAgents, the researchers fed it a dataset of 20 Chinese novels and their corresponding English translations. The system then learned to mimic the process of a human translation company, from initial translation to final proofreading.
In blind tests, professional translators and even GPT-4 Turbo itself preferred TransAgents’ translations to those of human translators and other AI models. While the system’s translations were not word-for-word replicas of the originals, they effectively conveyed the original author’s meaning, style, and cultural context.
While TransAgents represents a significant advancement in machine translation, the researchers acknowledge that there is still room for improvement. They are now working on developing new metrics to evaluate the quality of literary translations, as traditional measures like BLEU may not fully capture the nuances of this complex task.
This breakthrough could have a profound impact on the world of literature, making it easier for readers to access works from different cultures and languages.
More details: https://arxiv.org/abs/2405.11804
Demo here: https://www.transagents.ai/