Neural machine translation (NMT) is an algorithm used to translate words from one language to another. It is said that high quality NMT can determine the context of the translation and use models to offer a more accurate translation. Google Translate, DeepL, Yandex, Baidu Translate are well-known examples of NMT available to the public via browser automatic translation.
Let’s get deeper into it.
What is a “neural network” for translations?
Neural Networks are called this way because the system is vaguely inspired by the biological neural networks that constitute the human brain. Elements that represent the neurons are organized into layers (neural network).
- The network is constituted by units called “nodes” (like human neurons) and all the units are connected
- Each connection is a number that is called an “edge”
- Neurons and edge connection is represented by a specific number called a “weight”
During the processing the weight is adjusted, for example once the machine has more translated data to analyze.
Applied to translation, unlike the traditional phrase-based translation system, which consists of many small subcomponents that are tuned separately, neural machine translation attempts to build and train a large neural network that reads a sentence and the context.
How does the translation model work?
The simple answer is via a complex mathematical formula (represented as a neural network). This formula takes in a string of numbers as inputs and outputs a resulting string of numbers. The parameters of this neural network are created and refined via training the network with millions of sentence pairs (e.g. English and Chinese sentence pair translations).
Neural Machine Translation is using models
A machine can also be trained to meet the specific needs of a sector (legal translation, medical translation, etc.) or a customer’s area of activity, which will have its vocabulary.
NMT are primarily dependent on the training data used to train the neural network, as it learns to mimic the data it has been trained with. Many high-accuracy industry specific and custom developed machine translation (MT) models still incorporate both neural and statistical methods today to squeeze the best performance for our clients
Each sentence pair modifies the neural network model slightly as it runs through each sentence pair using an algorithm called back-propagation.
Translation model illustration
Neural Machine Translation can learn how to translate better
As opposed to some old engines still on the market (statistical and rule-based), a neural engine models the entire process of machine translation through a unique artificial neural network. Google starts using its own GMNT back in 2016.
Today, the machine – like the human brain – is capable of producing a reliable translation as well as learning a language, and is therefore constantly improving the quality of the translated material. To improve the performance of the machine, it is “trained” by human translators.
Google neural machine translation illustration
In practical terms, the NMT quality depends on
- A huge volume of translated data (words, segments of sentences, and texts already translated)
- The quality of the translation provided to improve the reliability and sophistication of the results
- The power of the computers used to create the models and also in some case to handle real-time translations
The most popular generic neural machine translation engines
Machine translation becomes better and better every year, but depend on training data and the algorithm quality. Depending on your needs, the machine translation engines can be generic or custom (with custom translation models). There’s no doubt that generic machine translation engines will reach the same quality as custom machine translation engines in the future, but that’s not yet the case.
Let see who are the popular generic neural machine translation engines
Google Translate
Generally considered one of the leading machine translation engine, they almost invented it 🙂
No doubt that, based on our latest tests, the quality is better than competitor and the number of languages is above competitors too.
DeepL
DeepL is a smaller company from Germany, originally based on the data of Linguee website. Quality is really close to the Google API, but they have fewer languages available.
Microsoft Translator
This is the translation tool you probably use in Office Word. This is another cloud-based neural engine, called Microsoft Translator. If you’re looking for an average quality but with small fees, this is the way to go.
Amazon Translate
Amazon Translate is also neural-based and is closely integrated with Amazon Web Services (AWS). The quality is poor compared to the 2 above, except in certain Asian languages.
Neural translation quality by language
The neural machine translation has increased its quality a lot since 2016. For example, in a major language pair like the first one in the table below, from 80% to 97% in 2022! And no doubt that in a couple of years it’ll be hard to spot a difference between a human translator and a machine translation in this language pair.
| NMT Translation | Human Translation | Quality | |
| English-Spanish | 5,428 | 5,550 | 97% |
| English-French | 5,295 | 5,496 | 96% |
| English-Chinese | 4,594 | 4,987 | 92% |
| Spanish-English | 5,187 | 5,372 | 96% |
| French-English | 5,343 | 5,404 | 98% |
| Chinese-English | 4,263 | 4,636 | 92% |
Some use cases for Neural Machine Translation
Ecommerce and business online websites
Translation opens a new market, this is simple as that. If you know your audience or analyze your audience with a statistic tool you’ll easily find a new potential country to target. Usually, it’s country that does not understand your language at all or countries with high demand for your product.
For example, in Spain the English language level is lower than in Norway, so an English translation might bring more conversion. Check this post is you want to know more about audience analysis.
Short timeframe to translate large amounts of web content
The neural machine translation models require a large amount of high-quality translated data to improve its neural networks, so it can rapidly produce precise translations.
A good example of the need to translate large amounts of content is to translate instructions to use a complex product like an industrial machine. This is typically a PDF of 1 thousand pages. With a neural machine translation, it could be done in a minute in multiple language.
Customer support ticket
The direct customer support is often the first contact with customer, so it’s crucial to give accurate information to resolve issues as fast as possible. Allowing content to be translated in his native language is a remarkable experience in that regard instead of the “we do support in English only” 🙂
Neural Machine Translation for repetitive content
NMT is especially effective at translations that is very repetitive, such as translation for similar products.
For example, if you’re selling 50 different headsets models online, using human translator only will require to do 50x very similar translation for similar characteristic of each product.
Benefits from Neural Automatic Translation for your website
Connecting Linguise translation with your website will bring all the benefits from the automatic neural translation to your customers, mainly:
- 80+ languages available
- High accuracy of the translation, especially in popular languages (around 90% of a human translator, up to 95%)
- Instant translation of all the website, no matter the platform or the content size
- Translation of highly repetitive content
- All the neural translations can be edited (see the video below)