However, Ng believes agentic machine translation has “huge potential for improving over traditional neural machine translation”. Limited testing using BLEU shows that the software is, to quote Ng, “sometimes competitive with, and sometimes worse than, commercial providers”.
The demo uses a “reflection workflow”. A large language model — in this case, GPT-4 turbo — is asked by an agent to translate text and then to reflect on the output. The model is asked to come up with constructive suggestions to refine the translation, and, subsequently, to apply them.
The idea is that an iterative approach using AI agents (software that performs tasks on behalf of a user) can potentially produce better results than a single prompt of a large language model (LLM).
Slator 2024 Language Industry Market Report — Language AI Edition
The 140-page flagship report features in-depth market analysis, language AI opportunities, survey results, and much more.
Other advantages? The system is “highly steerable”. Ng explained. “Simply by changing the prompt, you can specify the tone, regional variation and ensure consistent translation of terms (by providing a glossary),” he said.
These types of LLM applications are not exactly new for translation management systems and AI-enabled machine translation providers. And one in 10 translators already use LLMs to apply glossaries to achieve consistent translation.
However, it is not the use of LLMs per se, but the potential to use AI agents to leverage LLMs in an iterative workflow for translation — and beyond — that Ng wants to explore.
“We think this is just a starting point for agentic translations, and that this is a promising direction for translation, with significant headroom for further improvement,” he said.
Ng’s demo comes on the back of a paper from Chinese tech giant Tencent and Monash University that explored how AI agents can simulate the distinct human roles in a translation agency to produce and improve translations.
