Translation using Large Language Models (LLMs)

Large Language Models (LLMs) are one of the most exciting and disruptive technologies to emerge from the Generative AI era, providing new possibilities and opportunities for language service providers like CSOFT and their clients. LLMs are a relatively new development in the AGI space, with the deep learning architecture that forms their foundation only debuting in a 2017 Google white paper. That architecture is known as a transformer network, a pattern-identifying neural network that can scan and predict text thanks to two key attributes. The first is positional encoding, which means that the data in a text string does not need to be inserted in order, and the second is self-attention, which assigns values of importance to the words the transformers process. Much like transistors on a processor, Large Language Models emerge from the layering of these networks into a comprehensive and fast whole, all working together to achieve a common predictive goal, and with AGIs like GPT-4 growing in popularity and in skill, it is an essential area for companies to invest resources in.

This means translations can be done at a blisteringly fast speed with a reasonably high degree of accuracy. Yet, much like a race car is only as good as its driver, an LLM is only as good as its user left to their own devices, LLMs can be prone to hallucinations – strings of random text processed as a kind of error – and misses many of the nuances that exist in culture and common parlance. To prevent these types of mistakes, CSOFT uses the best of traditional translation technology – our human linguists and subject-matter experts – combined with these new AI tools to deliver the most accurate translations possible in the shortest time necessary.