Is there no point in segments?
In the late 1980s, with the development of computer technology, programs began to be developed to assist translators in their work. The former enabled the latter to work faster, deliver more coherent translation texts, and to make use of previously translated content. Today, advances in technology have brought the translation industry solutions based on artificial intelligence, which make it possible to create translations even faster, and sometimes instantly, an of a quality similar to the results of human work. Do the latest developments make it easier or do they preclude working with what are called CATs? I will try to answer this question below.
What’s the deal with CATs?
CAT is an abbreviation for computer assisted translation, and CAT software is the one I mentioned at the very beginning of the article. With CAT programs – and there are at least some of these on the market now – the translator, first of all, does not have to work in a conventional word processor. In a CAT program, the text to be translated, called the source text, is divided into segments, which in most cases are sentences. Each segment is suitably short, specially highlighted and positioned one below the other on the left hand side of the computer screen with a space on the right hand side for the translator to type the translation text. This ensures that the translator’s eye does not get lost in the strings of characters that make up entire paragraphs or pages. The features available in a CAT program will also prevent the translator from skipping a sentence, which is more likely to happen when working in a word processor.
All such segments or sentences in the source language paired with their translations are saved to a database used by the CAT program, which is called a translation memory. Such saved source-target pairs can later be used repeatedly when translating other texts, even if they are not 100% identical to the new content from those other texts. The likelihood of reusing saved translations increases if the translator works, for example, for the same client or with texts from the same field.
This would briefly be the basic idea behind CAT programs. In addition to this, they cater to many other needs of the translator by offering, for example, glossaries of terms, extensive autocorrection and proofreading and filtering options.
How AI works needs no explanation
While the average John Doe has never been confronted with CAT programs – as these are primarily designed to assist translation specialists who have a command of foreign languages – it is more likely that he has already used the various AI-based translation tools available and did not need any special explanation or foreign language understanding to do so. It is enough, for example, that he knew the address of a website where he could use a machine translation service, pasted the text he wanted to translate into the appropriate field, and after a while he could already read and further copy the translation text. And all that even in the fee-free version of the service.
Translation tools based on artificial intelligence have of course always been well-known to translators as well. They, however, speaking both the language of the source text and the language of the target text, often could not be satisfied with the first version of the translation text offered by the machine. Over the past few years, the quality of these machine translations has been steadily improving. Nevertheless, they still need to be revised by a human, and especially in those situations where the purpose of the translation is not just to generally convey the meaning of a text.
Two intelligences working together
Translators, working with machine translation tools available online, had to modify the still unperfect translation texts, and they were able to do so right at the time they were using the service. In other words, they did not have to, and still do not need to, correct the translation text separately in the word processor at a later stage of their work.
In one of the most well-known online services offering AI-supported translation, the said modification looks like this: the translator can place the cursor in front of the piece of text they would like to change or correct, and then other possible versions of the translation at that place will be displayed to them. All they have to do is select and click on the one they consider most appropriate. If no such one is displayed to them, they can also type it in themselves. And what is most important here: once the translator has made such a single change, it will automatically be applied to further parts of the text. The machine, in a way, cooperates with the human and adapts the entire translation text to the human’s choices, which significantly speeds up the creation of the translation text. With this kind of machine translation, typing is reduced to a minimum or even to zero, and thanks to the fact that the machine follows the translator’s choices, the resulting translation text gains consistency.
AI & CAT
Translators did not have to wait long to be able to use AI solutions in CAT tools. Machine translation providers have been offering additional modules to extend the capabilities of CAT programs for several years now. The plug-ins, as they are the thing in question, insert the ready-made translation text in the place where the translator would have to type the text themselves. The exception, of course, will be those situations where the translation text is retrieved by the CAT program from the previously mentioned translation memory – then there is no machine translation or typing on your own – but these situations will still not be common.
Plug-ins greatly speed up the translator’s work in CAT programs, reducing the time that is spent typing. The quality of the translation text they offer is also constantly improving, which by no means implies that the translator no longer needs to modify the suggested text. And it is in the context of this human modification of the machine-translated text in the CAT program that a question can be raised whether there is any point in having segments.
Each segment is a different story
By using AI-based translation tools available online, the translator can quickly and effortlessly correct the translation text as a whole in cooperation with the machine, which adapts to the translator’s choices. Machine translation plug-ins used in CAT tools, on the other hand, cannot treat the text as a whole, because they are “asked” each time to translate only one segment, which is usually a sentence. So there is no way an intelligent machine can correct the content beyond that one sentence, because it is not available to it at the moment. Moreover, the translator has to correct the translation text themselves with each successive segment, which is not the case when using, for example, an online machine translation service. They always have to type any such correction on their own, because in the editor of the CAT program they cannot just click and select another word from a drop-down list.
What about still having segments then?
One of the simplest conclusions you could come to having read the above would be that translation text is created in a faster, more convenient and simpler way when using AI-based translation tools. The text is created by clicking and making choices, without any unnecessary typing. In contrast, working with a text input into a CAT program, where it is segmented and cannot be treated as a whole by an intelligent machine, no longer seems so smooth, efficient and fast right now. Trying to answer the question asked in the title of the article, you might therefore be tempted to say that there is no longer much point in using segmentation in CAT programs in the face of the potential offered by AI.
If I were mostly to translate so-called ordinary and non-specialized texts on a daily basis, such as this article, for example, I would fully agree with the above statement. The content and sentences in an article for a company blog are unlikely to be repeated in any similar texts that I will translate in the future. I also do not need any dedicated dictionary of terms or advanced proofreading options to translate an article, for example, to check whether I have omitted any numerical data.
Things are different in the case of specialized texts, especially patent descriptions, which I normally translate at JWP – they tend to have content that is repetitive, and to work on their translation I need everything that I wouldn’t typically need in order to translate a simple article. So to ensure a translation text that is of high quality and consistent not only across and within its own body but also across the body of all texts in the field, I will go for the combination of CAT+AI, rather than AI alone, because of the additional features of CAT programs. The benefit of the latter outweighs the benefit of faster translation using AI only.
Will it someday be possible to bypass the problem that AI cannot treat a text as a whole in CAT programs? At the time of writing this article, I have not found any technological solution that addresses this problem, but with the current pace of development of new technologies, I think we will see it sooner rather than later.
