5 ways to use AI in translation for greater productivity

For some, AI-powered translation technologies such as neural machine translation (NMT) and large language models (LLMs) spell doom for human translators. For us, it's the opposite. Not only do we believe that there are reasons to be optimistic about the future of human translation, but we believe that the use of AI in translation is a key productivity tool for human translators.

 

As one of the world's largest language service providers, we see the localization landscape shifting all the time. We see, for example, how a pandemic decreases demand for translation from the travel industry but increases demand from healthcare organizations. We see how the rise in remote working raises the profile of eLearning localization. Or how the growth of audiovisual content creates a shortage of translators in this tricky area of localization.

 

Overall, we see a continuing growth in demand for translation that doesn't look like reversing any time soon. The businesses we work with are translating a wider variety of content into more and more languages. One of the ways we continue to deliver effective, accurate translations while meeting client expectations for turnaround times and cost control is to embed productivity-enhancing tools into the process. Increasingly, AI is finding its way into this toolkit. Here are five of the most common ways it is doing so – or likely to do so in the near future.

As our blog about the future of human translation pointed out, we don't believe AI is about to replace translators. But we're already using NMT to accelerate the delivery of quality, accurate translations at scale. And LLMs are likely to follow, once security issues are resolved and processes developed for reliably tackling their translation weaknesses. Partnerships between AI translation tech and translators are the future, with translators compensating for classic deficiencies of AI in translation, such as hallucinations, bias or failure to understand humour or metaphorical language.

 

Not all AI translation tech is created equal, however. Using inadequately trained AI can actually slow the process down, giving the translator a mess to sort out rather than a largely accurate translation to polish. The importance of AI training cannot be overestimated if AI in translation is to realize its potential as a productivity-enhancing tool.

Whether it will be translated by machine or human, content sometimes needs a good deal of preparation before it can be translated at all. For example, if you have a video to translate and don't have the script (as ultimately filmed), then you'll need to turn the audio track into a text-based format for the translator (human or machine) to work with. Or if you receive content for translation as a scanned image, you'll need to extract the words from the image file. AI technologies such as automatic speech recognition (ASR) and optical character recognition (OCR) are playing an increasingly important role in getting content ready for translation, simplifying our ability to translate a wider range of content formats. As with AI-powered translation, the output quality of these tools is rapidly advancing, making them an integral part of the quest to deliver more accurate translations, faster.

When a translator has a question about the content they're working with, it can be surprisingly laborious to ask a question and get it answered. Often the person who has the answer – a subject expert or someone who knows the context well enough – is several steps removed from the translator, causing delays and increasing the chance of miscommunication. If AI can reduce this dependency on a human support chain, it could greatly improve both the quality and speed of translation. This would be a classic 'support bot' use case, with AI trained on relevant existing content, queries and reference material so that it can offer an answer without human intervention. Only the thorniest of questions might still need to be referred up the chain for human input, saving a lot of time.

AI and humans have such different strengths and weaknesses. AI's lack of real understanding is a key weakness that humans can compensate for. By contrast, machines are orders of magnitude better than humans at finding patterns in large volumes of data. Applied to error detection in translation, this has real potential for improving translation quality assurance. AI-powered QA would use the tirelessness, power and consistency of machines to combat the ways in which humans typically make errors.

Human knowledge and language preferences are always evolving. This is one reason we believe that AI translators won't replace human ones. On the other hand, the right kind of AI – specifically, LLMs – can be extremely helpful in updating linguistic assets following a big change such as a rebrand, a rollout of new required terminology, or the adoption of more inclusive language. The generative nature of LLMs can be a real strength in this context, making it a powerful tool for cleaning translation memories, adapting the tone or style of content, or replacing a particular word in all of its different forms in a given language. As with using LLMs for translation, though, there are still security and process issues to be dealt with before this can become a standard part of the translation process.

The five examples given here just the tip of the iceberg of use cases for AI in translation. Some are already almost standard, such as the use of NMT, ASR, OCR and other linguistic AI tools we've not even mentioned. Some, such as AI-powered updates, have been proven as a concept but aren't yet ready for business-as-usual rollout. At RWS, we're always actively exploring ways to embed AI into our language services to improve delivery, without compromising quality or security. And we've barely scratched the surface of the possible.