Good, fast, inexpensive – how MTPE facilitates high-quality translations at low cost
For a long time, the brusque motto in the translation industry was “Good, fast, inexpensive – choose two!”. Since the advent of neural machine translations, and especially since the intelligent combination of machine translation and human post-editing (MTPE), this motto has clearly lost ground. Companies can now benefit from high-quality translations at low cost. And all this at the push of a button – if you know how to do it right.
Machine translations with untrained machines is not new territory in the language industry, but for a long time they served more for the amusement of colleagues than for reliable translations into foreign languages. Phrases such as “God Save the Queen” quickly translated by machine made the rounds in companies and on the Internet, lulling the translation industry into the belief that generic machines were usually incapable of providing anything more than just a good laugh.
The situation was very different, though, with trained or domain-specific engines that were fed data for a specific subject area or company-specific data. However, anyone who ever participated in these training programs in the company knows that the phrase “translation at the push of a button” was way off the mark: the months of training, retraining, and high costs for on-premise solutions never bore and still do not bear any relation to the output quality and potential savings for many companies.
Normally, machine translation never used to be considered, particularly for companies with low translation volumes and little or no training material but a concurrent high demand on linguistic quality. But the breakthrough of neural machine translation (NMT) has suddenly turned the tide for everyone.
Machine translation and savings “from a standing start”
Whether DeepL, Systran, KantanMT, Globalese, or Tilde: In the field of NMT, a large number of providers are getting in on the act, making their machines – some generic, some specific – available to interested users. DeepL in particular has turned the market upside down with fluent translations and low initial expenditure. If translations are available at the push of a button, virtually free of charge and within seconds, why continue to pay money for human translations? “Fast” and “inexpensive” – see the title – would therefore be covered “from a standing start” with little outlay on your part. But what about quality? Does MT really have what it takes to replace human translations, as is already the case in some companies?
A close look at the results shows that the devil is usually in the details. “Schalten Sie das Gerät ab” has been translated as “Switch the device on”, which is wonderfully fluent to read but is quite simply wrong. With the English language, a large number of non-native speakers would be able to spot such mistakes, but in the Portuguese or Japanese languages, for example, non-native speakers are much less likely to spot them. In order to ensure that such errors are found, with fast, inexpensive, AND good results, the magic abbreviation is: MTPE. The intelligent combination of machine translation and human post-editing ensures that the machine output is thoroughly checked. The result should be as indistinguishable as possible from a human translation, while still retaining the time and cost advantages of the machine. And although post-editing is a manual step in the automated translation process, an average of 30% in savings can quickly mount up compared to the classic translation process, depending on the requirements.
Requirements for the successful use of machine translation
To ensure that the result of an MTPE project never costs more than a human translation, it is necessary to know in advance what is feasible. The recommendations of some providers and experts as to which texts are suitable for the use of generic machine translation sometimes read more like a search for a needle in a haystack: no nested, endless sentences, but also no short texts without context. Standardized texts as far as possible, but no contracts. No overly technical content, but also no colloquial language. Manuals yes, but only without references to interface texts. And of course: no marketing texts! Does the success of MT really depend on the type of text?
Practical experience shows: a definite no! No types of text can or should be categorically excluded. In addition to the pure text type, a number of other factors influence the result. For example:
- Subject field
- Language pair
- Engine used
- Specifications (style guides, terminology)
- Experience of the post-editor
So, what is important is not rigid exclusion criteria, but an individual feasibility analysis. For some texts that at first glance appeared unsuitable for MTPE, a closer examination or a short test run showed a clear potential for savings compared to human translation. One of the decisive factors here is the post-editors themselves: Whether or not you can actually benefit from the time and cost advantages of MTPE depends to a large extent on the post-editor’s experience, structured approach, and split-second decision-making.
At the push of a button and with intuition
Contrary to skepticism, and sometimes popular opinion, the machines do not take the work away from the translators, but are available as an additional tool in everyday translation work. Pre-translation can lead to higher productivity and thus a higher word output for the post-editor. Where it is otherwise only possible to translate 1,500 words per day, it may be possible to post-edit 4,000 words. The price per word for post-editing is correspondingly lower. However, if the payment at the end of the day is identical, but a significantly higher volume of words has been post-edited instead of translated, larger jobs can be completed in less time. This ties up fewer resources and allows post-editors to work for more customers, or to work on more jobs from the existing customer base. Post-editors therefore play a key role in the MTPE process, because without their expertise in the field of machine translation for specialist texts, nothing really works. Really good results depend on numerous factors:
- The post-editor’s knowledge
- Fair payment
- Precise specifications
- Regular feedback
Feedback is always welcome. Feedback to colleagues and superiors, to customers and suppliers – that is to say, 360° feedback wherever possible. As in most fields, the demand for feedback in the field of MT is always great. However: How much sense does it make for the post-editor to give feedback if it cannot be used to improve the generic machine? For the machine itself: none at all. For the MTPE process: a great deal. This is because targeted feedback from the post-editor reveals weaknesses in the source texts (text structure) or gaps in the terminology database. The feedback to the post-editor, however, is even more important than the feedback from the post-editor: Which segments should have been adjusted but weren’t, what still had to be corrected during the quality assurance process, and what are the resulting guidelines for future MTPE projects? As already stated, post-editors play a key role in the MTPE process, and therefore it is a good idea to provide them with training and bring them on board.
Machine translations and post-editing have rightly triggered a small revolution in the translation market. Practical experience shows that generic engines that are available at low cost are also suitable for everyday translation work. The success of the process depends on the right partner, who should be well-versed both technically and in the specialist field and who, above all, never loses sight of cost-effectiveness. Even though a large number of texts are suitable for MTPE, the costs and time involved must always be kept in mind in order to take advantage of the intelligent combination of artificial intelligence and the human mind.