FAQ’s on machine translation

We at COMLOGOS also have high quality standards for our machine translation. That is why a machine translation project with full post-editing is processed using SDL Trados Studio, SDL Trados GroupShare and our carefully tested, filled and trained machine translation engine. After the machine translation output is generated, an evaluated and experienced post-editor works on the translation. In doing so, he takes into consideration the common aspects for potential errors that machine translation entails. This means that right from the quotation stage the focus is on quality and all necessary factors are observed during project execution.

All post-editors go through a multi-stage evaluation process before they are allowed to edit your projects. They must be able to prove their qualifications (according to ISO 18587) and experience, and must undergo tests that enable us to assess the quality of their work. As a result, only evaluated and experienced native-speaking technical translators work for us as post-editors in the usual COMLOGOS quality.

COMLOGOS works with a carefully tested, filled and trained engine for neural machine translation (NMT). Your texts will be used exclusively for your translation and not for training the engine. Since the engine’s servers are located in the EU, all activities comply with the strict data protection regulations of the European Union.

“Engine” in the context of machine translation is to be understood in the figurative sense, as it actually is a computer program. This engine is responsible for the generation of the MT output, must be trained intensively and continuously fed with translations.

An immediate delivery after the order is placed is unfortunately not possible. It actually only takes a few seconds or minutes (depending on the length of the text) until the MT output is generated. However, other factors contribute to the fact that the processing time is somewhat longer. The project has to be created and processed, i.e. the files have to be analysed, the prices have to be calculated, enquiries/assignments have to be sent to the post-editors. Afterwards the output must be post-edited and then an internal check takes place in addition. Overall, however, the processing time is much shorter than for a project with human translation.

The amount of work involved in PE varies from project to project. But on average, you can expect 4,000 to 6,000 words per day. It should be noted that PE can be performed for a maximum of 4–5 hours a day (with a break), as the cognitive effort is much higher than, for example, in the case of a translation or revision.

The quality of the machine translation output depends heavily on the engine used. Accordingly, since the effort involved in post-editing cannot be estimated without knowledge of the engine, post-editing is unfortunately not possible for a text that was machine-translated by you. We only offer the complete package of machine translation including full post-editing.

In principle, a revision is also possible for post-edited machine translations. It should be noted, however, that machine translation is still not in conformity with ISO 17100 after a revision is performed. Only human translations including a revision meet the requirements of ISO 17100.

If you order both machine translation and human translation projects from us, two TMs are automatically created for each language combination: one TM for human translation and one TM for machine translation. This prevents machine translations and human translations from mixing in the TM.

If you order both machine translation and human translation projects from us, two TMs are automatically created for each language combination. Using only one TM is unfortunately not possible in this case, as machine translations are not in accordance with ISO 17100 and therefore entries in the TM that originate from machine translation may not be used for ISO-compliant translations.

 

Post-editing and revision are both carried out by qualified native-speaking technical translators. The difference is that post-editing is done for machine translations, whereas revision is done for human translations. This is why revision is also called the 4-eye principle.

 

Unfortunately, it is not possible to certify post-edited machine translations. If you need a certified translation, we recommend that you order the human translation including certification.

At the moment, we do not yet offer any customer-specific engines. However, this does not present a disadvantage for you since we use a TM – as we do with human translation. Your customer-specific translations are stored in it, which can be reused for future orders. Two TMs are even used for machine translation: your normal TM for human translation and the TM for machine translation. This ensures that your translations are consistent in style and terminology and meet your requirements. You can of course contribute to this by providing us with your company-specific terminology and any reference materials. Furthermore, we work according to the regular translator principle. This means that the same post-editor always works for you and is therefore familiar with your requirements.

No, we don’t offer that.

We only offer full post-editing, as full post-editing achieves almost the quality of a human translation, i.e. the post-edited texts are grammatically and terminologically correct, consistent and stylistically appropriate. Only in this way can we meet our quality standards at COMLOGOS.

The aim of light PE is that the translation is merely comprehensible or readable at the end. This means that after PE there will still be errors in the text. What makes a text comprehensible and how many errors it may contain is of course subjective. To prevent misunderstandings and ensure that everyone in the target group understands the text, we recommend full PE.

We have deliberately decided to use full PE also for internal texts, contrary to common practice. This means that we generally do not offer light PE. Light PE aims to provide information only, i.e. to make the text comprehensible. The MT output is only roughly improved, whereby the output is checked for completeness and serious grammar errors. However, stylistic and syntactical inconsistencies as well as inconsistent terminology and formatting errors are permitted. Consequently, these do not need to be corrected (unless this causes misunderstandings).

Since we are certified according to ISO 17100 and maintain a very high quality standard, we cannot reconcile delivering faulty texts with our quality standards. Apart from that, cost saving would be rather low.

The quality of machine translation with full post-editing should correspond to the quality of a human translation. This means that, just as with human translation, full PE ensures that the translation is correct in terms of spelling, grammar, terminology, formatting, etc. The post-editor also ensures consistency and an appropriate style. The main difference to human translation is that the style does not have to be perfect.

However, machine translations with full post-editing are not ISO 17100-compliant and do not comply with VDE specifications either. So if you need a certified translation, we recommend a human translation including revision according to ISO 17100.

It can generally be assumed that – with intensive training of the engine – the linguistic quality of the NMT output is much higher than that of statistical and rule-based MT.

In rule-based MT, the system is based on grammatical rules for each language pair and a dictionary. These rules are supposed to represent the grammar knowledge of the speakers of the respective language. The engine analyses each sentence in the source language, tagging the words and phrases to highlight the grammatical components. The engine then searches its dictionary for translations for the tagged components and uses the coded grammatical rules to combine them into a sentence in the target language. Due to this approach, the output is often translated word for word and is therefore not very fluid or context-sensitive.

In statistical MT, the engine translates by analysing large volumes of translated texts and thus the relationship between source and target language data. In this process, each sentence is broken down into its words or word sequences (n-grams) and the associated translated sentences are analysed, identifying the most likely translation from several possibilities for each n-gram.

We currently do not offer pre-editing, as it can be very time-consuming under certain circumstances. Due to the costs involved, machine translation would no longer give you a (large) price advantage over human translation.

Anyway, we test in advance whether the respective text is suitable for MT. However, if it is not suitable, we can advise you on the topic of “MT-oriented writing” so that you are able to prepare your texts for MT in the best possible way.

The internal check is a final quality assurance measure that is included in the price and is generally carried out for every project before delivery. It is therefore not possible to skip the internal check.