Machine translation (MT) is becoming an increasingly integral part of modern translation programs. The rapidly advancing development of machine translation technology—and the amount companies are investing—support this. Also, MT quality is getting better and better: the results have become increasingly fluent and accurate. An important milestone was the 2016 presentation of the "new" Google Translator based on neural machine translation (NMT). It was a quantum leap for machine translation, so to speak. In 2017, DeepL, another provider of a free neural MT solution, appeared, and even more MT providers have risen up since then.
The differences in quality were rather small among the new players, but they have improved with continuous engine training. However, there are still significant differences in the number of language pairs available. Google offers far more language pairs than DeepL. Only Microsoft Translator can currently compete in this space, though SDL Language Cloud currently offers approximately 3,000 language combinations for users to access.
1. What is post-editing?
It’s important to pause and consider the task of post-editing (PE). PE differs considerably from classic translation. It is actually a special form of review.
A translator always aims to deliver a translated file that is as faithful as possible to the original, but a post-editor works differently. There are generally two types of post-editing: light post-editing and full post-editing.
In light post-editing, the focus is on ensuring that the target text correctly reflects the meaning of the source, so spelling or grammatical errors are of secondary importance. Even style is often overlooked unless the errors distort the meaning of the source. The most important consideration is whether the translation provides the reader with sufficient information.
In full post-editing, the post-editor produces a translation that comes as close as possible to the same quality level as that of a human translation.
Here is a list of the things a post-editor would correct when performing full post-editing:
- Formatting/tagging: ensuring that all tags used in the source text (e.g., for text formatting, cross-references, etc.) are present in the target text and in the correct locations.
- Terminology: checking for appropriate use of industry- and client-specific terminology as well as correct use of proper names.
- Fluency: checking and establishing readability to make the text as fluent and similar to human translation as possible.
- Country specifics: adapting country-specific currencies, units of measurement, formats for dates, numbers, addresses and punctuation, cultural references and more.
- Style: adapting the style for the target audience and to client- and project-specific requirements.
2. The three MT strategies
To better understand the practical application of PE, it’s important to explain the scenarios of MT and TM applications that produce the “starting point” for the post-editor. When an enterprise deploys MT, there are several strategies that can be used:
- Complete document pre-translation. The complete document is pre-translated. This means that all target segments up to a defined fuzzy match threshold are reused from the translation memory. All other segments are machine-translated. This way, the entire file is populated with translations from either the TM or the MT engine. From there, a post-editor edits the entire document, bringing it up to the required quality level.
- Pre-translation into a translation memory. The post-editor receives a translation memory with the raw translations from the MT engine. The translator still has access to an array of fuzzy matches that they can use, if applicable, to create the final output. The advantage of this method is that the post-editor only edits the pre-populated segments when needed and can use the fuzzy matches from the translation memory if they want.
- Interactive translation with on-the-fly MT support. As the post-editor types a translation, suggestions from the MT engine are offered for use. This gives the post-editor the greatest possible freedom in terms of wording because they are not starting with a populated response that they have to edit. Instead, they have more choices and control over the language used.
3. The evolution from translation to post-editing
There are plenty of translators who are skeptical about machine translation and post-editing. Some fear for their jobs; others dislike the nature of post-editing work. But there is really no reason to fear. First, the use of machine translation is not suitable for all types of text. For texts that have “emotional weight”, idioms, puns, humour or complex content, MT can be so poor that a post-editor must fix everything—in which case, MT has no value. Machines can substitute words and groups of words from one language to another based on patterns taught to them through training, but they cannot handle all the aspects and complexities of human language. In addition, if there are errors or ambiguities in the source content, machines lack the judgment required to understand them and still produce a correct translation.
Post-editing will therefore not replace classic translation, as there are still cases where it’s better for a professional translator to start from scratch. It is simply an additional service that complements existing offerings.
However, the acceptance of and willingness to work with MT technology is important for success in this field. When translators are expected to learn how to post-edit, it is absolutely necessary to explain to them how machine translation works, where it can help and when it shouldn’t be used. Explain to them the nature of the errors that can occur. Also, provide them with guidelines on the levels of quality so they can calibrate their work accordingly. In addition, post-editors should be able to work in their familiar environment; i.e., they should be able to use the same tools that they use for classic translation.
All this said, training alone does not make someone a great post-editor. Post-editing skills are acquired through daily practice. Our experience shows that even the linguist with thousands of words of post-editing experience may be a problematic post-editor if the experience was acquired through very random tasks and over a very long period of time. Also, linguists whose experience with MT is dated and lies years in the past might not have the right understanding of the actual quality of today’s MT outputs.
In summary, translators can learn post-editing as an additional offering and skillset that makes them more marketable and available for more types of work.
4. How to become a post-editor
Becoming a post-editor takes more than the usual webinars that are offered free of charge by many companies. Some language service providers offer in-house trainings for their employees or freelance translators. Also, some universities offer seminars on the topic.
Critical content includes:
- A technical introduction to MT;
- The history of machine translation;
- Possible applications of MT and the associated decision criteria;
- The differences between post-editing for statistical-based MT (SMT) and neural MT (NMT); and
- Practical post-editing exercises in both light and full post-editing techniques.
That last point, practical exercises, is very important. These give the prospective post-editor the opportunity to ask questions and learn the nuances between the two types of post-editing first-hand.
In any case, training or other continuing education to become a legal post-editor is an investment in both the career of the translator and in the future of enterprise MT development and deployment. The increase in the use of MT will lead to a greater demand for legal post-editing, and there’s no going back.
At RWS Alpha, legal post-editing is what we do. If you are interested in setting up machine translation for legal content, contact us to discuss the levels of quality that can be achieved by post-editing to find the best-fit process for your needs. And, check out our post-editing ebook, A Comprehensive Guide to Legal Post-editing Programs.