Terminology management made easier with large language models

In recent years, the use of AI technology has brought significant changes to the translation industry. One of the most notable examples is OpenAI's GPT model. This model is getting a lot of attention because it promises to produce coherent and contextually relevant translations.

 

But did you know that this same technology could be applied to terminology management, specifically to perform complex tasks that are traditionally reserved for human linguists?

 

Our Linguistic team tested this hypothesis using large language models (LLMs), such as ChatGPT and text-davinci-003. They conducted experiments to determine whether LLMs can accurately distinguish syntactically and semantically ambiguous terms and replace them with new terms.

 

The team ran a term replacement experiment using three terms that are ambiguous between different parts of speech (e.g. build as a verb or as a noun). The results were very promising. GPT was able to differentiate and target specific parts of speech entries, replacing the verbs with the new term, while leaving the nouns untouched. GPT even correctly adjusted the new term to the structure where it was added, ensuring that the correct syntax was preserved. For instance, the past tense form “built” was replaced by past tense form “created” and not by the present form “create”. GPT also successfully recognized the targeted terms even when they were distorted by typos or deliberately introduced language errors.

 

See example table below. Our prompt was: In the following sentences, replace all occurrences of the word "build" by the word "create" but only when "build" is used as a verb. Do not change the word "build" when it is used as a noun.

The experiment also tested GPT's ability to disambiguate semantically ambiguous words. It produced excellent results, opening the door to new applications such as adaptation from a given language to variants of the language.

This technology can potentially transform the dreaded term change implementation into a much quicker task, helping you keep your translation memories always up to date with the latest project terminology. This can be achieved without the need to invest in hundreds of human hours.

 

However, implementing LLM-driven term change alignment may not be trivial. Different models perform differently, and the success of the term replacement tasks varies depending on the strategy used.

Despite the potential benefits of LLMs, the expertise of human linguists will still be needed. Their work will be augmented by prompt design tasks and testing learning strategies to extract the best AI performance.

 

LLMs have the potential to revolutionize terminology management. While they may not offer the ultimate no more tears formula, they can help make this complex task more efficient and less labour-intensive. The use of LLMs should be viewed as a supplement to human linguists, rather than a replacement.