What is adaptive AI?
Adaptive AI is used in countless areas. However, one thing is always true: it is not static but is capable of learning and constantly evolving in a data-driven way. The more data is available for it to draw on, the better (more on this below).
In the context of translation, we understand adaptive AI to mean a translation model that develops based on its learning material, i.e. existing translations. This process can be repeated as often as desired with new information, which further improves the model. This training makes it possible to specialize an AI translation system down to the last detail for a specialist field and specific wording – for example, with your company data or with sources that match the domain and text type.
What types of specialization are there?
Broadly speaking, translation models can be divided into three levels of specialization:
Generic AI translation
e.g. Google Translate, Bing Translator, DeepL Free, Textshuttle Free
The non-specific “out-of-the-box” variant. This model was trained with a wide range of data. As a result, it can process all types of texts in a comprehensible way, but often falls short of specialist needs and is often unable to recognize context correctly.
Let’s say you need to translate a text about the Swiss pension system from German to English. “BVG-Obligatorium“ is a well-established term in this field and describes the mandatory contributions according to the “Occupational Pensions Act (OPA)”. However, as the Generic AI translation model used in this case has not been specifically trained for this subtlety, the rather nonsensical “BVG compulsory” is used.
AI translation with light customization
e.g. DeepL Pro, Textshuttle Pro
Most fee-based systems allow fixed terms to be stored in glossaries. In this way, you can reliably avoid misunderstandings and mistranslations of individual words.
Using the glossary, you can now specify that “BVG-Obligatorium” should always be translated as “mandatory OPA”.
AI translation with data-driven customization
e.g. Textshuttle Enterprise
The fully adapted version with specific AI training goes one (big) step further. We specialize a translation model using existing texts from your company. This aligns the output with your corporate language in terms of style, formatting, sentence structure and frequently used terms. In addition, the correct translation of ambiguous passages and terms such as “split option” can be deduced much more reliably from the context.
Example: Textshuttle Enterprise
In our example, Textshuttle Enterprise learned from the training data and automatically optimized the sentence structures to frequently used phrases and the overall context.
With adaptive AI, the text can be translated much more precisely for your specific use case. This reduces the need for corrections and thus makes the process more efficient and cost-effective.
But what does the learning process look like in detail?
How does adaptive AI for translation work?
The path to a fully specialized translation model (in the sense of data-driven customization) involves three steps, which can be repeated in cycles:
Train
First, we work with you to determine the database for the initial round of training. Two principles apply here:
1. The more, the better
As a rule, training will produce meaningful results starting from100,000 segments per language pair. That’s equivalent to around 3,000 A4 pages of translated text.
2. The higher the quality, the greater the precision
Your previous translations from the translation memory database provide the best basis, as they already fully match your language and content. If not enough material is available, we can consult online databases with training material from your specific area of expertise.
Our AI experts collate this data and filter it to ensure it is clean. We thus ensure, for example, that the collected pairs of translated sentences are available in all relevant languages and are clearly equivalent to each other.
Then the initial training begins. The model searches for patterns in the data it has been provided and recalibrates itself based on its findings in order to align the output as precisely as possible with the training data.
Use
The first version of your specialist AI translation solution is ready and goes live. Professional translators (e.g. in-house language services or translation agencies) play a particularly important role here. They use the system as a basis for their work and revise the AI output with post-editing.
The results – now verified as free of errors – are automatically returned to your translation memory database. This database of professional translations forms the raw material for further model improvements, meaning translators and the engine support each other.
Refine
The AI Research team evaluates and validates the newly translated data and regularly refines the translation model. This means that your corporate language is always kept up to date and translation quality constantly improves.
When is adaptive AI worthwhile for you?
The advantage of a specialized AI translation solution is clear: it ensures more accurate translations. But what does that mean in concrete terms?
For professional translators, a specialized solution demonstrably reduces post-processing work by up to 50%. And in fully automated processes (when translating thousands of product texts for your online store, for example), it significantly reduces the risk of embarrassing mistakes.
Specialized systems are therefore particularly worthwhile for companies that have to translate many publications professionally and/or manage large volumes of text fully automatically.
Are your time and budget limited, but your standards for translation quality high? Then a tailor-made solution might be right for you. Let’s talk.
Cover image via envato elements