Translation and technology

Machine translation (MT) is a translation performed by a trained machine without human intervention. A fundamental distinction must be made between rule-based or statistical MT and neural machine translation (NMT). The latter has become the leading MT technology in recent years thanks to enormous advances in the development of artificial intelligence (AI) – keyword ‘machine learning’. We also use it in our company to ensure that you, our customers, benefit from its advantages.

NMT is based on a computer programme that is structurally similar to biological neural networks. These ‘neural’ networks of the MÜ are trained through an initial intensive training process in order to achieve a relatively high degree of reliability in translation and autonomy in learning. This applies both to ready-to-use generic MT engines (such as Google Translate, DeepL, etc.) and to adaptive ones, which can be trained even more specifically in terms of a subject area or style, for example. For such training, the learning machines require large, high-quality bilingual text corpora, which enable the neural networks to identify patterns in human translation. From these patterns, they derive abstract word representations in the form of vectors, which are then used in the translation process.

i) Enormous cost savings, especially when post-editing is not required

ii) Extremely fast translation speed, i.e. reduced turnaround time

i) Large bilingual, high-quality text corpora are required for NMT to approach the level of human translation.

ii) Furthermore, the quality of NMT depends on the language pair and direction. Certain language pairs (such as German-English) are ideally suited to machine translation due to their structural similarity, while others (such as Arabic-English) are less suitable or even unsuitable, at least for the time being.

iii) Another criterion is the type of text and its complexity. Certain types of text with manageable syntactic and semantic relationships – such as news texts – do not pose a major obstacle for neural machine translation. However, as the degree of complexity increases, as in demanding legal or technical texts, for example, it becomes more difficult to achieve a high level of quality using machine translation alone. For marketing and advertising texts, to give another example, NMT is simply unsuitable. This is because these texts require creative cultural transfer, a feel for linguistic subtleties and the free play of words, which machines are not capable of despite AI.

i) One of the biggest pitfalls of neural machine translation is the appearance of a fluent text, even though the content has been translated incorrectly. This appearance can be so convincing that the MT system leaves individual words or entire expressions from the source text untranslated in order to maintain the flow of the target text.

ii) Another shortcoming of NMT, at least to date, is the lack of or insufficient text coherence and terminological consistency. Although the MT system is capable of translating an isolated sentence correctly, it is not able, unlike humans, to perform the higher-level task of establishing relationships between several sentences and the terms they contain, thereby producing a coherent text with consistent terminology.

It is virtually impossible to give a general answer to this question, as too many factors need to be considered and the final MT solution is always tailored to the individual customer. Once the necessary requirements for machine translation have been met and the quality level of the target text has been defined, a pilot phase lasting several weeks is usually carried out, during which the new text material is fed into the MT system for training purposes. Only then can a reliable assessment be made as to whether pure MT is also economically viable. If you would like in-depth advice on this topic, please do not hesitate to contact us.

We generally recommend machine translation – even generic machine translation – for internal company communication, for example to gain an initial understanding of the content of an email, i.e. for pure information gathering (gisting).

Even if pure NMÜ is not worthwhile for you, you can be sure that we will use this technology not comprehensively and fully automatically, but selectively and in a controlled manner to your financial advantage. The advantage of our integrative approach can be illustrated with an analogy to ‘autonomous driving’: just as the driver in a car that drives ‘autonomously’ thanks to artificial intelligence still keeps his hands on the steering wheel so that he can intervene in an emergency, it is the professional translator who, as it were, can steer the translation process at all times from the cockpit of his CAT systems. Consider the seriousness of a translation produced solely by a machine with just one tiny but serious error that could result in property damage, legal liability, damage to reputation, economic loss or damage to health. Our translators ensure that such damage does not occur.