As generative AI starts to mature, its powers and limitations are becoming clearer. Tools like ChatGPT are making an impression as translation solutions, putting new capabilities in the hands of language professionals. So let’s take a look at how language professionals are using the power of generative AI and integrating it with other translation tools.
Earlier this year, we tested ChatGPT at a range of language tasks to see how good the latest generative AI tools are at translation. We found that ChatGPT produced similar quality to Google Translate and other AI-powered translation tools. In some cases, ChatGPT performed slightly better than Google Translate and in others, it performed slightly worse.
This is impressive, considering that generative AI tools aren’t designed specifically for translation.
Similarly, generative AI tools can achieve tasks that translation tools like Google Translate are incapable of. For example, users can ask ChatGPT to first translate a piece of content and then ask it to modify the translation after spotting some linguistic inconsistencies in the first output.
So with generative AI tools, you have the ability to provide feedback and make follow-up requests. This helps to refine the translated output further, correct mistranslations and produce different results with each prompt.
The most obvious way for professional translators to use generative AI tools is to mimic the traditional machine translation process. All this requires is a prompt like “translate this text into Spanish: [text]” or “translate this into Greek:[text]”.
In many cases, this will produce results comparable to using a tool like Google Translate. For example compare the translated output generated from ChatGPT vs Google Translate for the following English Twitter post into Greek : “Police on Naxos arrest six for unauthorized beachfront occupancy” (Source: Kathimerini English Edition). We can see that ChatGPT’s output is better due to the correct nuances it has applied to its translation, but both versions could still benefit from human post-editing for optimal results.
Output from from ChatGPT.
Output from Google Translate.
As mentioned earlier, our tests have found ChatGPT to slightly outperform Google Translate with some Spanish translations and slightly underperform in other areas depending on the language.
What does this really mean though?
Well, it’s very interesting to see that ChatGPT quite often does a better job at translating conversational texts like social media posts. This makes sense to some extent, considering the fact that GPT-4 has access ‘to a larger and more diverse dataset of 1 petabyte (compared to its previous version), which includes web texts, books, news articles, social media posts, code snippets, and more’.
This also explains why ChatGPT occasionally interprets and translates casual language more effectively.
The differences between ChatGPT’s and Google’s output lead us to the next role generative AI can play: comparison. Language professionals can compare the results of generative AI and machine translation tools to improve the overall quality of results through post-editing. With time, and as these technologies improve further, they’ll also learn which types of translation tasks each tool is more adept at and where their relative weaknesses lie, helping them to choose the best tool (or ideally a combination of tools) for the completion of certain translation tasks.
If you want to discuss the impressive capabilities of generative AI technology and how to implement such tools into your language projects, get in touch with us to discuss our Generative AI+ solution.