As you have learned, getting a good translation depends heavily on the translator’s ability to understand and extract meaning from the original text. These days, texts can be translated by people or by computers (for example, Google Translate, DeepL Translator, Microsoft Translator or ChatGPT), or sometimes by a combination of the two. But there are some significant differences between the way that people approach translation and the way that computers approach translation. Knowing more about the strengths and limitations of both people and computers can help you to better understand the risks involved in using automatic translation tools. In addition, knowing more about how people and computers translate can help you to prepare your texts in a more translation friendly way in order to increase the chance of getting a good quality translation.

Professional translators: they get it!

Before a translator can re-express a text in another language, they first have to understand what the original text is saying. Only once the meaning of the original text is clear can the translator can begin the process of finding the best way to express that same message in another language. If a translator cannot understand the original text, then they cannot translate it effectively.

Because translators are highly trained professional communicators, they can sometimes puzzle out the intended meaning of a poorly constructed text before beginning to translate it. For instance, translators can often spot typos, recognize faulty grammar, research ambiguous meanings, and compensate for a variety of other types of issues in the original text. In other words, a professional translator would realize that the original sentence “Let’s eat kids!” is probably missing a comma and would know how to translate and punctuate the sentence correctly in the target language (i.e., “À table, les enfants!”). Of course, in the unlikely event that the text happens to be about cannibalism, then a professional translator would know that the comma is not missing and they would translate accordingly (for example, “Mangeons des enfants”). Human translators are capable of understanding meaning, and they are able to use common sense, contextual cues, and real-world knowledge to help them to interpret the most likely meaning of a text, even when it is not expressed particularly well. But what about computers?

Automatic translation tools: pattern matching and counting

Free online automatic translation tools, such as Google Translate, DeepL Translator or Microsoft Translator, are becoming increasingly popular. Although these tools cannot replace professional translators in all circumstances, they can be helpful for meeting some types of translation needs. However, to get the most out of these free online tools, it pays to be careful about what you put into them. Remember: garbage in, garbage out!

Unlike people, computers do not actually understand language. They process it, but they don’t understand it. So how do automatic translation tools work? The current approach to automatic translation is a data-driven approach known as neural machine translation or NMT. To make these data-driven tools work, they require data. In the case of translation, the data consist of texts that have already been translated by professional translators.

Computer scientists begin by gathering a VERY large collection of existing texts. This is called a training corpus. In the training corpus, they align the texts sentence by sentence, so that it is clear which sentences in the two languages are equivalent. For instance, here is a short excerpt that illustrates some aligned sentences that have been taken from the bilingual website of the University of Ottawa. The English and French texts have been aligned at the sentence level.

Of course, this excerpt shows just a handful of sentences taken from one website, but to train an automatic translation tool effectively, a training corpus needs to have millions and millions of aligned sentences originating from many different sources.

Neural machine translation tools are based on artificial intelligence (AI), which means that the tool contains an artificial neural network. The emphasis on artificial is important here. People possess a real neural network (i.e., their brain) and are capable of truly understanding a text. In contrast, AI-based tools do not understand anything. Instead, they can only process texts using techniques such as pattern matching and counting. In this way, NMT tools can partially imitate what translators do, but not fully. NMT tools do not understand the underlying meaning of the texts; they only look for superficial similarities.

A term that is often associated with AI is machine learning, but the learning is also artificial. In the case of translation, a computer scientist will feed the very large training corpus of aligned sentences into the automatic translation tool, and the NMT tool will look for patterns. For example, it might identify the pattern “the University of Ottawa” in the English sentences, and then identify a corresponding pattern – “l’Université d’Ottawa” – in the French sentences. The computer doesn’t understand the meaning of a text. It doesn’t know what a university is or what a city is. It only sees patterns of characters, and it can count how often these occur. So the computer can keep track of the fact that it has seen multiple examples of the pattern “the University of Ottawa”, and it can note that whenever that pattern occurs in English, there is a corresponding pattern in French (i.e., “l’Université d’Ottawa”).

People can often learn new things by seeing just a few examples, but remember that computers are not actually learning. They are just looking for patterns. In order to be sure that what they are seeing is actually a pattern and not just an accidental combination of words, they need to see many, many examples. This is why the training corpus for an automatic translation tool needs to have millions of texts. This is also why automatic translation tools make mistakes: they don’t actually understand the text.

Unlike a person, a computer won’t try to figure out what you meant to say. It will simply process what you actually said. A computer won’t compensate for typos, poor grammar or other language issues. It will accept the problematic input and translate it “as is”, which in all likelihood will result in a nonsensical translation. Garbage in, garbage out! Here are some translation that were generated by Google Translate. Because the English texts contain an error or can be interpreted in more than one way, the translations proposed by the automatic translation tool ended up being incorrect. Do you think that a human translator would have made these mistakes?

Choosing between professional translators and computers

For translation projects, there are three competing factors to be taken into consideration: time, cost, and quality . Of course, in an ideal world, you could get a high-quality translation in a short time for a low cost. But in reality, it can be difficult to achieve all three at once, which is why this problem is known as the triple constraint. Usually, you can get two out of three, but you may need to compromise on the third one. For example, in the case of translation, you can have a translation that is:

• Fast and cheap, but of lower quality;
• Fast and good, but more expensive;
• Good and cheap, but it takes longer.

Automatic translation tools are very fast, but they produce lower quality texts. Meanwhile, human translators are very good, but they take a bit longer. On the surface, online automatic translation tools seem to be cheap, but if the mistakes that they make have serious consequences, then this choice could end up being costly.

A key thing to take into account when choosing whether to use an automatic translation is tool is the consequence of getting a poor translation. If the translation is for your own personal use or for internal use within your team or company, then the stakes are probably lower than if you plan to share the text widely or publish it on your company website. Likewise, if the content of the text is about planning the office holiday party or organizing a company softball team, the consequences of a poor translation will be relatively minor. In contrast, if the text is about a financial audit or a new product, then a low quality translation could have serious repercussions, such as tarnishing the company’s reputation or even creating legal problems. Of course, there is a third option also, which is to ensure that a text that has been translated initially using an automatic translation tool will later be verified and, if necessary, corrected by a competent person.

In any case, regardless of whether your text will be translated by a person or by an automatic translation tool, it pays to write in a clear and translation friendly way. Even though human translators are more capable of spotting and correcting or compensating for problems in the original text (for example, typos, homophones, missing punctuation), a text that contains such problems will certainly slow a translator down, which could end up delaying or increasing the cost of a project. The upcoming sections will present some tips that you can use to create translation friendly texts.