machine translation

ChatGPT was launched in November 2022, and it feels like the world – or at least the future of information work – has not been the same since.

ChatGPT is the first service that makes an artificial intelligence based on a large language model (LLM) available to the everyday user and provides an easy user interface for it. Even in this case, artificial intelligence isn’t an actual “intelligence” but a mathematical model that calculates which words are likely to follow one another. However, ChatGPT does this more smoothly and with better versatility than the previous models.

At best, ChatGPT improves human productivity

At first, it may seem that the model can do anything: it types code, writes snappy articles and responds to questions. Naturally, it also does translation! In ChatGPT’s user interface, it is possible to request translations and other texts using natural language prompts in the same way as when asking from another person. I asked it for a poem on selling localisation services, and it even managed to do that. In terms of style, the output was much better than what I myself would have been able to produce.

However, when you look at texts produced by ChatGPT a little more closely, you will see that it suffers from the same issues that plague machine translations by proficient neural networks, for example – the translations are often correct but usually only almost correct, and the AI may come up with sources that it cites as if scientific research.

Due to the errors it makes, ChatGPT is, in a way, very human. Regardless, this does not make the quality of its text poor in general but rather reminds us of its limitations. If you want to make use of the output of an LLM, it would be a good idea to have a human check it or, at the very least, you should acquaint yourself with what kind of text it produces in a specific context.

If you remember and are mindful of these limitations, ChatGPT is certainly a useful tool in various industries. When used properly, it increases human productivity by coming up with ideas, checking and making drafts.

What about translation?

ChatGPT is surprisingly adept at translation. In the chat user interface, translation is easy to do by using free-form requests. What separates ChatGPT from traditional machine translation tools is that you can also give it instructions for the translation task. Below are a couple examples of this. In the first, I requested a translation in the style of a pirate ( what could be more useful?), and in the second, a couple of alternative translations.

And it can also translate into several languages of your choosing with just one prompt:

So, producing translations is easy, and they seem fairly correct as well (while my language skills aren’t sufficient for assessing all of the examples). In some cases, it even adds comments that explain its decision-making process.

Thus far, there have only been preliminary evaluations on the accuracy of the translations produced by ChatGPT compared to the prevailing neural machine translations (NMT), and it appears that the translation performance of LLMs cannot compete with the more specialised machine translation tools. The difference may not be huge, but it is still significant enough to make a difference in professional settings. However, there have not been many comparative studies yet, so we should not be too quick to draw further conclusions. In any case, it seems that ChatGPT beats the neural network models in specific context recognition or memory tasks.

LLMs in professional use

The entire translation industry has now turned its gaze to large language models, wondering if they will replace NMT-based machine translation tools or even translators themselves. Talk of the eventual redundancy of translators is nothing new, but that hasn’t happened so far – and I don’t think it will happen now. But LLMs will undoubtedly have some sort of impact.

The examples below are simple and easy, but even they show that you get lots of possibilities for translation by using different prompts compared to traditional machine translation tools. The current machine translation tools always produce one version in a set way. You cannot influence their translation output except by training the tools on materials of your choosing or by using an adaptive system and feeding it with texts in advance.

Through the use of prompts, ChatGPT can make adjustments on the fly, which you could make use of with the help of example texts or translation memories, for example. In the same way, it also offers possibilities for editing the ready text to follow a certain style.

As ChatGPT supports well the production of entire texts, it may shift part of the translation work away from traditional translation, thus moving the focus more on creating a new text in another language. Of course, this is what translation basically is already, but due to practical reasons, translations are often very closely tied to their source language. In time, machine-assisted methods used in the translation services industry could more clearly be divided by their intended use. This means the method more closely linked to the source language could be retained in technical texts but would give way in more creative translations.

Despite the hurdles, there are endless possibilities

LLMs will inevitably become part of additional services in translation. For example, in term base work, ChatGPT can already be successfully used for rough work on short texts. Perhaps there are more uses in text analysis as well? There must be many more uses that I do not even realise or which are not even a thing yet.

So, there may be endless possibilities, but there are also many hurdles. At the moment, one of the most significant hindrances is data security: ChatGPT cannot be used privately, which prevents its use in many situations. However, I expect services with better data security and privacy to find their way to the market soon. When it comes to translation tools, accuracy also plays a crucial role, and as long as LLMs cannot compete with neural network-based models, they will have less use in professional translation.

Currently, there are more questions than answers regarding the use of LLMs in the language services industry, and no one can say for sure where these models will be used. In the coming years, LLMs and ChatGPT are certain to shape the language services industry in a decisive way and create new ways of working and new competence areas, starting from what kinds of prompts users know to use when requesting texts from AIs.

My guess is that we will have new important LLM-based machine translation tools and text-production solutions, the mastery of which will (for a time) become a competitive edge. And even this tool is used by humans, whose productivity it may increase a lot.

Nearly everyone has heard of machine translation. Google Translate is an important and frequently used tool for many, and machine translations can also often be found on social media platforms. Almost every widely used social media platform gives the user the opportunity to see posts machine-translated into their own language. Machine translations can be found in many places, and the user may not always even realise that a translation is produced by a machine.

Machine translation methods

Although machine translation hasn’t been used extensively until quite recently, it has been an objective of data processing since the 1940s. The method of producing machine translations has changed a number of times over the years, sometimes basing the process on certain rules, at other times on statistics, and so on. These approaches, which are still partly used, had their advantages but have since become quite marginal. The biggest public progress in the field of machine translation in recent years was made in 2014–2015 when the first scientific article on neural machine translations was published and neural networks were used for the first time in a machine translation competition. At the moment, the market is completely dominated by neural machine translations.

Neural machine translations, often abbreviated as NMT, are the most important AI or, more accurately, machine learning application in the language industry. Neural networks calculate vectors for words that also take into account the context in which the word is used. As a result, neural networks very rarely produce grammatically incorrect text. There are some mistakes, of course, but they are primarily not grammatical and relate especially to semantics and terminology.

Today’s machine translation tools produce high-quality text, but they cannot replace human translation expertise and understanding of the broader context. Instead, they complement human competence and provide a basis for a number of new services.

Where to use machine translation

According to a Finnish proverb, fire is a good farmhand, but a bad master. The same applies to machine translation. Machine translation can be used to provide certain services that would not be possible without it. However, trying to use machine translation tools to produce all the translations you need may end up with you shooting yourself in the foot.

Let me now highlight a couple of applications that are ideal for using machine translation. (These are not the only possibilities.)

Support sites

The support sites of web services can grow quite large over time or be very extensive from the start. New companies in particular rarely have an unlimited budget for localising their support site, and even large companies do not want to unnecessarily spend huge amounts in this area.

Machine translation is an invaluable tool for localising customer support websites. In principle, it provides almost instant translations into numerous languages at a very low cost. I say “in principle” because you have to remember the limitations of machine translation. The use of machine translation alone does not guarantee a flawless end result and, in the worst case scenario, the user may receive completely inaccurate information and misuse the product or service. Where can you find good guidelines for using machine translation? Perhaps the most important piece of advice is to know your product and be aware of its special requirements. In other words, if you are manufacturing, say, circular saws, do not rely on machine translation for your support site if you want to make sure that your customers don’t cut off their fingers. If your product is not directly hazardous if misused, you can be more flexible.

Even on support sites, it is not advisable to use the same processes for all content and languages. Through analytics, you’ll probably know which articles are used the most and which issues lead to the most support requests. It is probably a good idea to translate them more carefully than pages that are used very rarely.

The best toolbox for localising support sites contains various tools: translators, translation memories and machine translation tools – not to mention glossaries and other resources. Translation memories are used to teach the machine translation tool essential content and to achieve the highest possible quality in the raw translation. Translators edit the most important pages – sometimes all of them – to ensure perfect translation quality. Your language service provider will help you manage this complex package.

By the way, here’s an important but often forgotten tip: if you offer your users a “raw” machine translation, let them know. They will be more likely to forgive minor mistakes and they can check any ambiguous parts against the original language version, even if they are not fluent with it. If a user mistakes machine-translated materials for a human translation, they may become annoyed and consider the translator and, therefore, the service provider incompetent.

Internal materials

Internal materials do not always need to be translated perfectly. The translation should simply be very close to the original text so that a reader who is not proficient in the original language can understand its content. These texts include, for example, invitations to tender or comments from external users on various services. Such translations may have such a short lifespan and a very limited readership. As internal users, readers can also be very forgiving about small mistakes.

When internal materials are needed quickly, people can be easily tempted to use free online services where text can be copied and pasted. This is often a workable solution, but it has some obvious shortcomings: document formatting is lost, the results are not based on previous translations, and data protection concerning your information is at risk.

Fortunately, you can pick the best parts of each approach. Document-based machine translation considers the context, and the process can take advantage of the organisation’s approved materials through translation memories and adaptive machine translation tools.

An automated and fast solution does not have to be the simplest solution. If necessary, this process can also be integrated with a certain degree of a human touch, depending on the subject matter and purpose of the text.

Benefits of a great machine translation partner for your organisation

There are many different ways and purposes of using machine translations. Therefore, the delivery method of machine translation must also be flexible. Through API, the customer can receive machine translations directly to almost any system and, through various integrations, there are ready-made ways to deliver texts across multiple systems, such as content management systems and online stores. An ordering system based on entire documents is an easy way to make machine translations available to even more occasional users in your organisation.

The quality of machine translation is neither a random factor nor a “standard” that could not be influenced. A knowledgeable machine translation partner makes use of the customer’s text resources to improve the results, instead of simply offering what any customer would receive in any case.

A great machine translation partner also offers the most important building block: language expertise. Together with the customer, the machine translation partner is able to find the most suitable combination of machine power and human expertise.

machine translation

Will ChatGPT revolutionise machine translation?