Why AI isn’t always the best solution

How machine translation tools work

Machine translation tools are developed using multilingual databases and trained to match segments or phrases according to meaning. In neural machine translation (NMT), this is done using what are called neural networks—computer processing systems designed to mimic the way the human brain works, at least to the extent that that is possible. This allows NMT to go beyond simple segment matching and take the context of the translated document into account, resulting in a more accurate translation and more appropriate terminology.

However, it’s important to understand that NMT systems simply process and reproduce whatever is fed into them. This means that the quality of the translations they produce is directly proportional to the quality of the databases they’re fed. That’s why the more specialized a document is, the less reliable the NMT output is likely to be. NMT systems also cannot tell when terminology needs to be consistent, which can create problems in texts with defined vocabulary or multiple synonyms.

There are more specialized NMT systems that can be trained through usage and include functions for correcting output, but they require a significant investment and are not available to the general public—certainly not for free.

Neural machine translation and generative AI: What’s the difference?

Although NMT and generative AI both run on large language models and are based on neural networks, the way they process data is different. Simply put, NMT processes a document sentence by sentence, while generative AI “understands” the original document and generates an equivalent document in the specified language. Research conducted by experts in the field shows that translations produced by generative AI are actually less accurate than those produced by NMT.

Confidentiality and cybersecurity

NMT improves itself by learning from what users enter into it for translation. The terms and conditions of some of the most popular apps even state outright that any content fed into the system will be reused. This is an issue from three standpoints: quality, confidentiality and cybersecurity.

In terms of quality, because NMT trains itself on its own output, which can be of dubious quality from the outset, the quality will diminish over time.

With regard to confidentiality, it’s important to keep in mind that all information fed into these systems is automatically captured and reused in one way or another. Although it won’t re-emerge in its entirety, it can sometimes be easily recognizable. There has already been at least one case where a confidential contract translated using mass-market NMT caused a breach of confidentiality serious enough to derail a major financial transaction.

As for cybersecurity, any confidential information entered into these systems becomes vulnerable to cyberattacks. NMT servers are often located in countries other than Canada, so information security is also a concern. If your company handles personal information or other confidential data, it undoubtedly requires that said data be kept on servers located in Canada. Using mass-market NMT automatically violates that rule.

When and how to use NMT

The NMT available online is ideal for understanding the general content of a document or quickly translating a short piece of writing on a general topic intended for a limited audience. It often produces very good results for general or lightly specialized documents.

However, there are serious risks to using NMT for documents that are sensitive or confidential in nature or that pertain to specialized fields requiring specific language and terminology. In those cases, it’s preferable to work with a professional translator.

Reasons to work with professional translators

Professional translators are familiar with the tools available on the market and know how to use them properly. Most importantly, they know how to verify NMT output to produce an accurate translation that complies with the codes, terminology and language of the field in question. Professional language service providers:

• Invest in dedicated translation software that can be improved and “taught” how to assess corrections to the raw output
• Populate their NMT databases with quality-controlled translations
• Create glossaries to ensure their translations are consistent and adapted to each client’s specific terminology
• Know how to advise their clients on the most appropriate process depending on the document and the situation
• Engage their professional civil liability
• Comply with a code of ethics and best practices

Editing NMT isn’t as efficient as it sounds

Professional translators are increasingly being asked to edit texts translated using NMT. But this is not always to the client’s advantage.

Depending on the complexity of the topic, the quality of the writing, and the NMT system used, the document can take much longer to edit and ultimately end up costing more than hiring a professional translator in the first place. As the old saying goes, you get what you pay for.