The development of DeepL: From Linguee to industry leader
Founded in 2009 in Germany under the name Linguee, DeepL originally began as an online dictionary with the ambitious goal of developing a neural machine translation system that would surpass the quality of traditional statistical machine translation.
The engineers at DeepL utilised the latest deep learning techniques and trained their models with the extensive data from the Linguee database.
With success: since 2017, DeepL has enjoyed immense popularity, with over one billion users benefiting from its services. Today, it supports 28 languages and offers an impressive 650 possible translation combinations.
DeepL provides its users with various options, including free and paid versions, as well as the choice between a web interface and a standalone translator. The free version is suitable for personal use, while the paid version includes advanced features for businesses.
In addition to translation services, DeepL has also made a name for itself in the field of artificial intelligence, particularly text generation. Its latest project, DeepL Write, launched in early 2023, aims to establish itself as an English writing assistant and outperform competitors such as Grammarly.
How does DeepL work?
Not all machine translation solutions work the same, and over time, systems have become increasingly sophisticated. Until 2016, machine translation systems were based either on rules that were created manually or on statistical methods based on multilingual corpora (large amounts of parallel texts). These methods translated based on words or phrases and searched for statistical patterns.
Nowadays, most common machine translation systems use neural networks. This is known as neural machine translation, a form of end-to-end learning in which the program’s neural network considers the entire input sentence at each step to generate the output sentence, rather than just considering a few words on each side of the translated term.
DeepL is another example of such an NMT (Neural Machine Translation) system. Thanks to deep learning algorithms, it generates translations that are more human-like than those of statistical MT machines. The network architecture also makes it possible to learn from large amounts of data and adapt to new contexts.
What can DeepL do?
As previously mentioned, the core strength of DeepL lies in its NMT system, which enables it to produce more accurate and natural-sounding translations compared to traditional statistical MT methods. Here are some of its outstanding capabilities:
Multiple language support: DeepL supports translations between a variety of languages, including English, Spanish, French, German, Chinese, Japanese and many more. Users can translate text seamlessly between different language pairs without having to use separate tools.
Context awareness: DeepL’s AI engine can understand the context of the input text to ensure that translations are more contextually accurate and retain the intended meaning. This is particularly useful when working with idiomatic expressions and complex sentences.
Translation quality: DeepL is known for the quality of its translations in certain language pairs, and this is highly valued by users. Its ability to capture context and deliver accurate translations is praised by professionals and language enthusiasts alike.
Translator plugin and desktop applications: DeepL offers browser plugins that integrate with various applications, including Microsoft Office. This enables direct text translation without having to leave the program currently being used. Alternatively, the desktop application, which is compatible with the most common operating systems, can be downloaded from DeepL.
Functionality for the translation of documents: In addition to text translation, DeepL allows its users to upload entire documents for translation. This is a useful feature for individuals and organisations working with multilingual content.
Data protection: Based in Germany, DeepL is subject to the data protection laws of the European Union. Although the company does not disclose many details about how data is handled in DeepL’s free machine translation, it emphasises that all DeepL Pro subscribers can enjoy a secure and encrypted connection for all translations and data is not stored on their servers once the translation is complete.
API integration: DeepL offers an API that allows developers to integrate its translation services into their applications, websites or services. This ensures a seamless translation experience for end users.
Evaluation of the accuracy and performance of DeepL
Determining the accuracy of a machine translation system can be extremely difficult. Results vary depending on the language pairs involved, and also on the types of text being translated – after all, a highly technical financial document is not the same as a blog post.
In addition, accuracy is relatively defined due to the dynamic nature of language and unmeasurable factors such as the author’s intended meaning. For example, expectations for accuracy of grammar and punctuation for a novel differ from those for a social media post.
Nevertheless, most users report that DeepL tends to be quite accurate, especially for European language pairs. DeepL’s own experiments support this claim. However, the way the company presents the data could potentially be seen as biased. A more reliable indicator of accuracy is therefore user feedback, which is overwhelmingly positive, especially in terms of:
- interpretation and equivalence of idioms and slang
- naturalness
- register-sensitive translations.
Ultimately, each user will have to make their own decision after testing DeepL in their own particular applications. The good news is that the free version allows you to try it out for yourself free of charge.
Advantages and disadvantages of DeepL
Neural machine translation systems such as DeepL open up a wide range of possibilities. As one of the leading providers of NMT technology, DeepL has the following advantages:
Pretty accurate: It can understand the broader context of words and phrases to produce more accurate and fluent translations and continuously improve.
Able to learn: Neural networks can be trained quickly through automated processes, as opposed to the costly and labour-intensive manual methods required for conventional machine translation.
Flexible and easy to integrate: Via APIs, extensions and plugins, you can integrate it into most browsers or software and apply it to many file formats.
Customisable: You can usually customise the output of DeepL through terminology databases, brand-specific glossaries and other data sources to improve results.
Cost-efficient: Like any other NMT engine, DeepL enables highly accurate and fast translations at a fraction of the cost.
Scalable: As your translation needs grow, DeepL can help you easily handle the increased demand.
However, while the development of DeepL has made great strides, hardly any machine translation system will ever be perfect – and DeepL still has some way to go before it can match human-like accuracy.
How much DeepL will improve in the coming years remains to be seen and is difficult to predict. At the moment, it still lags behind human-like capabilities, such as:
- being able to ask questions
- understanding context
- identifying errors in the source text
- recognising irony
- translating creatively
- making considered decisions
- doing research
- keeping an eye on consistency
- selectively omitting or adding information
- adding glossaries or annotations.
While the ability of machine translation applications to handle these types of tasks remains limited, manual post-editing is likely to remain an essential part of any translation workflow in the near future.
Using DeepL for professional translations: Important tips
When using DeepL for machine translation, it is important to focus on use cases where you can get the most out of it without compromising quality. These include:
Low-visibility or low-traffic texts, such as internal memos, website footers, emails, social media posts, internal analyses etc.
Repetitive technical texts that only need to be fit for purpose, such as user manuals. However, where possible, a human translator should post-edit this content to avoid potential user experience issues.
Low-risk texts that do not need to be 100% accurate or comply with complex regulations – a good example is user-generated content such as product reviews, where consumers generally do not expect high quality.
Quickly generated text, such as chat or email support messages, customer enquiries, etc.
Large volumes of text with fast turnaround, such as hundreds of product descriptions that need to be published quickly.
Frequently changed content – think feature and information updates.
Most of these use cases require only light manual editing (a post-edit) to ensure accuracy and clarity. You may even be able to work with the raw output if the content is not business-critical.
Some other types of content that require higher accuracy require more time-intensive post-editing. These include:
Product headings: Highly informative and concise, they often contain proper nouns and ambiguous words, and their word order is usually relatively free, which can lead to ambiguity.
Translations between language pairs with different syntax: Japanese and Spanish, for example, require reordering of words and phrases, making well-formed sentences more challenging for machine translation engines.
Content that influences the customer experience: This includes information databases, FAQs, warnings etc. These have to be as accurate as possible to achieve the goal of efficient customer support.
SEO meta information for the backend: Some SEO elements, such as image alternative texts and captions, have low visibility but need technical optimisation, e.g. targeted keywords in the target language, for improved search placement.
Conclusion
Overall, the article provides an informative overview of DeepL, its functionalities and areas of application, as well as essential insights for the effective use of this AI translation service.
