Machine translation is like a very sophisticated weapon; you should know how to use it.

Post-editing is a process resulting in the optimization of the translated text and the delivery of a reliable and satisfactory service. At Intertranslations, we have a team of experienced linguists, translators, and language solutions specialists who draw on different theoretical backgrounds and focus on the latest trends that dominate the translation industry. We bridge the gap between translation and technology by capitalizing on our know-how in translation, linguistics, and technology. One of our main interests is in MT tools and post-editing, and we understand the importance and benefits of post-editing, as well as the unique role of the post-editor. We also have extensive knowledge of machine translation systems and can provide detailed explanations of how they work.

What is the profile and role of the post-editor?

Based on research, we have approached the role of post-editor as an upgraded role rather than two different roles, that of translator and post-editor. In the survey question “What is the ideal post-editor profile for you?” to which language service providers, translation companies, and translators responded, we notice that they all agree that we need a degree in translation, post-editing training, and post-editing experience of more than 3 years.

Thus, we consider post-editing as the translator’s upgrading, upskilling, and not as something completely different. We understand, of course, that it is something very specialized and needs training and specialization either in translation or in a particular field.

We would like to show you how we understand the skills of a post-editor. This understanding does not come from research or entail figures; it comes from the bibliography we have studied.

We have linguistic skills, language skills, textual competence, cultural competence, and knowledge of the domain, the subject. More specifically, one needs to know both the source and the target language, communicate competently in at least two languages, have cultural and intercultural competence, as well as competence in the subject area. There is also technical competence, i.e. knowledge of machine translation systems, terminology management skills, quality assessment skills relating to the body of the text — in some cases of statistical or example-based machine translation tools — and skills in editing the controlled language. Also something else — optional, admittedly, but with an eye to the future: programming skills to create macros and run an automated error correction.

Furthermore, we have the basic skills, such as psychological competence, which allows the post-editor to handle subjectivity and apply post-editing standards, as well as strategic competence, which helps post-editors make decisions and choose between alternatives without style-related concerns.

What are the differences between a post-editor and a translator?

We need to define these roles somehow and see the evolution of the post-editor in relation to the translator. Some skills, competencies, and things that translators should be aware of are common. An example is that they should know both the source and target language. They must also be specialized in a field to be faster, more productive, and efficient in terms of quality.

In the beginning, a translator does not need to have specific knowledge of text editing, terminology and its management. On the contrary, a post-editor should have very sophisticated, advanced knowledge of text editing, be able to use macros and coded dictionaries, and manage terminology at a more advanced level to be productive. He/She should also be familiar with CAT tools, translation memories, machine translation technology, the types of post-editing, and the various levels of expected quality.

The post-editor must have an advanced knowledge of text linguistics and it is important to have technical knowledge — not so much cultural knowledge, which in general is essential for a translator.

The translator pays more attention to detail. In the translation process, he/she must pay a lot of attention to every detail, and he/she has the time to do so. On the other hand, the post-editor pays attention to speed, i.e. how to be faster, how to use his keyboard and mouse faster, and how to use whatever tools available to be faster. However, this does not mean that the post-editor should not pay attention to quality.

A translator, a traditional translator, may often not want to do machine translation. In that case, if he/she does it, he/she will not do it well: this is what our latest Memsource workshop and the research on this matter showed.

As for the role of the post-editor, it is also worth talking about pre-editing and the controlled language. It is useful, though not necessary, to use a controlled language and some programming skills may be useful for automated error correction purposes.

Can a reviewer be thought to have a similar role to that of a post-editor?

We need to define and differentiate the role of the post-editor from that of the reviewer. A review is seemingly when we proofread a text. This may lead us to believe that a reviewer can do post-editing very well. However, the two roles are completely different, because post-editing output errors are machine errors, while revision errors are human ones. The way these two work, the way the error is produced, is completely different. Therefore, the errors are made at different levels and derive from a different mental process. The post-editor must be aware of the machine errors and, thus, be able to predict them. Human errors are not as predictable, and we follow other processes in order to detect them.

What is the post-editing process?

Let’s have a look at the how-to, how the post-editing is done. We read the MT output, make a comparison with the source text and evaluate the quality in each segment based on the instructions. Then, we have to make appropriate decisions and process the text to improve it. So we have to read the source text, understand it, read the output, understand it, find the errors and then correct them. There are different ways to do all this; it is not a straight line. We can read the source first, read the target and start post-editing or vice versa; or it can all be done in parallel, that is, we read the target, read the source, and post-edit at the same time. The factors that can affect the quality and effort are the format of the file, the domain, the style, and some statistics regarding the document — for example, the length of the document, its complexity, whether it is of good quality, because some texts are not good at all.

What about the machine? What kind of machine translation systems are used and what “difficulties” can they create?

Let’s take a look at machine translation, specifically neural machine translation and the difficulties it entails. This type of machine translation is used today in 99% of the projects we handle, at least in translation companies. The downside of neural machine translation is that the translations are too fluent. The result seems to be fluent, but it is not perfect. We call it “masked,” meaning that the errors are hidden as if you have put a mask on them. They are not visible and are too difficult to detect. They are not errors at the stylistic, grammatical, or syntactic level. There is significant improvement in those regards, but while the translation looks perfect to you, there might be a term that is not correct. Because you do not have the necessary experience in a field, you might get carried away. You might think that something is correct, that it makes perfect sense, but if you are absent-minded even for a second and fail to do the necessary documentation research, it becomes a very serious mistake. To put it simply, this happens because what the machine does is understand the words that precede and follow, but this cannot work at the textual level. So the results produced are very good, but not perfect. This means that the more the machine and its results improve, the more necessary the translator and the post-editor, i.e. the human factor, become. We need highly specialized experts.

To post-edit or not to post-edit? When do we chose to post-edit and in what kinds of projects?

This depends on the content, the text, the machine translation system we have and, ultimately, the purpose of the text. If it is a text for publication, then it is not very appropriate to use only post-editing. We would suggest adding a revision step, always informing the client that we have used post-editing. There should be transparency on the part of everyone involved, from the client to the supplier. Only then is it right to use this service.

Some texts are not suitable for machine translation for two, or rather three, reasons: confidentiality, structure, or because they contain information that is critical for human life.

Confidentiality has to do with the security of the data given by the client, who has to know that it will be imported to a machine and whether we, as a translation company, have taken all the necessary measures to keep it confidential. We know that data is unsafe in the existing engines, google, etc.

The second reason relates to the structure of the text. If we have a legal text with very long sentences, according to the data available now, we know that the post-editing will not be good, because of the structure of such texts. It does not mean that we cannot do it; we can, but the result will not be as satisfactory as it would be for a technical text, usually containing instructions and sentences up to ten words, fifteen at most.

The third reason is the type of the text. If a text in the field of life sciences contains instructions for a medical device, which are critical (for example, there would be a big problem if you accidentally translated “do” as “do not”), post-editing would not be appropriate. Texts where a risk to human life is involved, even if they are about a machine, are not suitable, from an ethical perspective, for post-editing and machine translation.

Let’s talk about marketing texts. Based on the research performed, when machine translation is used for these texts, the output lacks creativity. The translations are not as creative as they would be if we translated the texts from scratch. So whether post-editing is appropriate or not is a matter of quality.

What are the styles/types of post-editing?

There is light post-editing and full post-editing. Light post-editing has to do with corrections that are strictly necessary, for example in terminology, accuracy, or fluency, so that the meaning and significance are conveyed properly. In this case we do not dwell on typos or anything else that does not affect the meaning, and it always depends on what we have agreed to do, in other words, what you can do in the time available. If you are asked to do light post-editing at 2,000 words per hour, obviously you will not dwell on typos, spelling, and syntax. However, you will try to make sure that there is nothing of key importance that can cause a critical error and damage because of inaccuracies and poor understanding of the text’s message.

Full post-editing is full post-editing at all levels, meaning that the result is comparable to a translation done by a translator.

We do not really agree with the distinction between light and full post-editing, because neural machine translation gives results so fluent that it does not make much sense to talk about light post-editing. One would make only few and targeted corrections at a deeper level. Light post-editing is still valid if we have a statistical machine translation system, while full post-editing would be suggested for neural machine translation. Light post-editing may suit a junior post-editor, but if we want to save resources, full post-editing would be better for an expert post-editor.

From which languages to Greek do we get better machine translation results?

So far, from our post-editing experience, we do not think the source language is particularly relevant. We have linguists translating from three languages into Greek. The result is just as good, but it depends on how the machine is “trained” and what type of machine is used. If we use neural machine translation, the result will be quite fluent. For Greek targets, the machine generally works well. Even free google translation works quite well. Greek does not have the difficulties that other languages, like northern or Asian ones, have. The results for those are not very good even when the machine is trained. It is, therefore, clear that it has to do with the level of training of the machine and the system used.

What about post-editing training? What problems and options do we have?

To work as a post-editor, it is important to be able to understand this service and that there are no gaps in the training, not only on the part of the post-editors themselves but also of the translation company. Universities face challenges as to how such training should be carried out, if it exists at all. The most important thing is to determine how all parties involved (post-editors, universities, translation companies) should work together for such trainings to be successful both on an academic and a practical level.

Concerning education, one can say that we lack relevant educational programs. To be more specific, we may find post-editing training at a university, but in a post-graduate course, at an advanced stage of studies, or it may not be global; that is, it may happen in Greece, but abroad post-editing may not even be heard of.

Another problem is the lack of clear instructions, as there is no specific set of instructions. If you work with several translation companies, each company or client may have different instructions. Obviously, it is too tiring for translators to have to read many different instructions when the context of the service is not well defined.

While in university, students should familiarize themselves with the theory and practice of post-editing and be able to identify machine errors if they want to do it in the future. It would also be useful to have some productivity tests in the university to test problem-solving and decision-making skills and find ways to develop these if someone wants to do post-editing after university.

The most important thing is cooperation between translation companies and universities. There could be some projects in the form of research for the university and actual ones for the company. In these projects, there should be a team with a supervisor from both the company and the university. This way, the university can identify mistakes and discuss them during the courses, and the company can allow students to develop through continuous involvement in a large project.

Currently, post-editing training is offered by one colleague to another while there could be an individual trainer. CAT tool providers, universities, translation companies, and some associations provide such training. You can always find information in books and online. The training is provided through workshops, job shadowing, internships, webinars, courses, such as those in universities, and mentored freelance work. All this is currently available.

Find out more information about MTPE here

Edited by Anna Dimkos, student at AUTH, translated by Lydia Kokkinidi