AI is Evolving the Translation Industry Rapidly
Everyone can see that AI is rapidly transforming many industries, and the language translation industry is at the forefront of this transformation. Converting one language into another programmatically has been a challenge AI has been working to solve since the initial development of AI technology. All of the AI-related breakthroughs in the past several years (Google’s development of NMT, LLM’s potentially replacing NMT, advancements in computer hardware, tech giants investing in the development of AI for translation) have accelerated the progression of AI translation dramatically. Whereas ten years ago we were warning against AI translation, now we are proactively encouraging AI integration (in the right situation) to many of our existing and prospective clients.
The quality of AI output is no longer in question, now it’s a matter of how to integrate it in a way that allows a customer to gain the benefits while mitigating the risks.
While we are big proponents of AI integrated translation workflows, we still firmly believe that in the majority of cases AI-Assisted Human Translation (AIHT) is the safest way to incorporate AI into the process. In this blog we’ll discuss what a basic AIHT workflow looks like, and how we can best integrate AI into the translation process.
How do we implement AI translation safely?
The basic answer to this question is that we keep humans in the process - or “Human-in-the-loop” as it’s known inside the industry. The traditional translation process is what’s known as a Translation / Edit process which involves two separate, professional, native-speaking human translators to create and edit the translation. A simplified traditional translation workflow looks like this:
In comparison, a simplified modern AIHT workflow looks like this:
You can see the shift in workflow from using two human translators to using AI to replace the initial translator with a human editor then editing the AI translation. This is possible now because the quality of AI translation has improved enough to make editing a realistic task. Prior to Google’s breakthrough in NMT the output of AI translation required too much editing, and was so prone to error in the form of bad translation passing through the edit process, that this workflow was not realistic.
Not a One-size-fits-all Solution
The picture painted above is a very simple illustration of how AI can be integrated into the translation process. This does not take into account the variety of different content types that may or may not be suitable for AI translation (chat content vs clinical trial documentation), or the target language (Spanish is a much better candidate for AI translation vs Japanese), or the myriad other challenges that may be faced during implementation (dealing with the design layer, reusing existing translation, etc.).
Integrating AI into your translation workflow requires a strategic plan to ensure that you are optimizing your ROI, while avoiding the potential risks of distributing incorrect, inaccurate, or potentially offensive or dangerous translations.
Here are some of the things to consider when planning for an AI integration:
Selecting the right AI translation engine - Which translation engine or LLM is the right one to use? In Intento’s “The State of Machine Translation 2024” report they evaluate 52 different MT Engines and LLMs. Is an MT engine or an LLM the right application for your specific content and language(s)? Intento’s guide can help you decide.
Application - Can you rely on AI across the board or do you need to apply it strategically? Our recommendation when initially integrating AI into your translation process is to not apply one solution across all content and all languages. For example, don’t just use Google Translate for all languages and all of your different materials (marketing communications, technical documents, software). Research the different engines available and their strengths and weaknesses and start with a pilot program for each content type and languages. Test, measure, and adjust. In some cases, like translating marketing material, you may find that you may not be able to apply AI because the benefits do not outweigh the challenges in application.
Existing resources - Once you’ve chosen the AI engines you’ll be using you’ll want to leverage any existing translation resources you have available. Glossaries, style guides, existing translation, can all be used to pre-train an AI engine to get better results right out of the gate.
Testing - Do you have a process set up to assess the quality of the output you’re receiving from the AI engine? Do you have internal teams that can review the output and provide feedback? If you do, you’ll want to set up a process for capturing that feedback and using it to improve the output of the AI engine through training.
Next Steps
We’ve provided a very simple example of how AI can be integrated into a modern translation workflow, along with some of the strategic considerations that you’ll want to be thinking about prior to integration. While it may seem complex or risky to consider an AI integration, we have seen that AI is at a state where it can safely, and believe it or not easily, be integrated into the majority of translation workflows for clients that are focused on gaining the benefits that AI provides, while being strategic about managing the risks. If you would like to learn more about how we build our AI workflows for our clients or more general information about AI please feel free to contact us.
If you’re interested in reading more you can read our blog entry “AI-Powered Translation Workflows Explained (Simply)”