To properly understand the role of neural networks in machine translation, we need to first take a step back in history and look at what came before: statistical machine translation (SMT). This translation system is based on massive amounts of translations, also called bilingual corpora. Using these corpora, the system calculates the statistical likelihood that a text element in the source text matches an element in the target text. This creates rules and patterns that are then used to translate new sentences. Drawbacks of SMT are that it struggles with context and requires multiple models for translations, each with its own design and specific training.

A (partial) solution to these problems is neural machine translation (NMT). First of all, this variant of machine translation is an so-called end-to-end system, meaning that only one model is required for translations. In short, and as the name suggests, NMT uses neural networks to generate translations. Word embeddings, or collections of words in the form of vectors, are crucial here. With neural machine translation, a neural network trains on large amounts of parallel texts, where each word in the source language is converted into a corresponding vector – a series of numbers representing the word. Through multiple layers of artificial neurons, the network then teaches itself to convert the source language vector into the target language vector. This target language vector is then read out and voila, your previously Dutch text can now be admired in France, China, or Japan.

The task of converting source text into a vector and reading out a target language vector to generate the actual translation is performed by an encoder-decoder neural network – currently the most commonly used architecture for neural machine translation. However, the first NMT systems struggled with the fact that longer sentences were often poorly translated. Long-term dependencies, the complex relationships between words not immediately adjacent to each other, proved difficult to translate. Luckily, the tech world is constantly finding a solution for every gap, and a solution has also been found for this problem: the self-attention mechanism.

The human in the model

When hearing the sentence a Lannister always pays his debts, series lovers immediately think of the hugely popular fantasy series Game of Thrones. They do this not because they have thought extensively about the meaning of each word in the sentence, but because a lightbulb went off at the word ‘Lannister’. The self-attention (or attention) mechanism tries to take advantage of this human cleverness. Using mathematical techniques, this mechanism helps the NMT system determine which input is most relevant in generating a response (a translation) and which is less so. As a result, longer-term dependencies are modeled, and the model works faster, better, and more efficiently.

Some benefits of machine translation, particularly NMT, have already been mentioned, but let’s outline the main ones for you:

• Accuracy: Research has shown that NMT systems such as Google Translate come very close to human translators for common (European) language pairs – think English-German and English-Spanish. Moreover, this variant of machine translation has the ability to ‘learn’ over time, making translations ever better. Where SMT systems previously struggled with context and sentences were broken down into separate segments, NMT works at the sentence level and takes context into account more effectively.
• Accessibility: As long as you have an internet connection, machine translation is accessible to everyone. And by everyone, we mean really (almost) everyone. Google Translate offers over a hundred different languages, and with DeepL, you have the choice of up to 800 language combinations.
• Speed: A professional translator delivers an average of 2,000 words per day, a translation machine does this in less than a minute. For a serious rush job, a machine translation can therefore be the way out for your project.
• Price: Many translation machines are free to use. So even with a small budget, a translation is to the end.
• Specialization: The properties of neural networks make NMT systems very suitable for designing models specifically for fields with their own jargon, such as the medical or legal world.

Perfect for e-commerce

NMT is particularly a handy tool in the world of e-commerce. In an online playing field that covers the entire world, companies want to reach as many people as possible and therefore translate their website into a large number of languages. Because: the more customers, the more revenue. But NMT is not only a way to increase the customer base in a cost-efficient manner; it also helps to improve the customer experience. With the help of machine translation, companies can quickly translate the feedback and reviews of their customers. This makes improving products and services much easier, and companies not only provide for a new target group but also for satisfied customers.

However, most e-commerce companies still do not fully trust machine translation. That perfect brand slogan with a pun that you’ve been pondering for weeks often loses its creativity when you have it translated by a program like Google Translate. A professional translator, on the other hand, knows how to handle this.

Sometimes the middle ground is the outcome, and in this case, that means post-editing. A language specialist, the post-editor, improves a text that has been translated by a machine. In this way, you combine the speed and low costs of machine translation with the quality of a professional translator. Post-editing can be a very suitable solution for texts that have been translated reasonably acceptably by the machine, but the keywords still need to be optimized. Post-editing is also a good option for texts that need to be translated regularly, such as annual reports or user manuals. A translation machine can handle this well, and a post-editor checks the texts to prevent translation fails.

Be aware that post-editing does not cost you more effort than it is worth. With texts with multiple storylines or a specific target group, a translation machine leaves little of the original text. Then the repair work takes so much time that hiring a translator is a better choice.

Simply Translate is the place for your machine translation Are you looking for cost-efficient translations and want to use the best technology for this? Then you’ve come to the right place at Simply Translate! Using our cutting-edge platform, we automatically handle incoming translation requests, ensure that your content is turned into a workable file and then deliver an automated translation. Of course, you can also opt for post-editing, in which case we engage one of our native translators before handing over the translation.

Want to know more about our machine translation, post-editing or the possibility of training your own datasets? Then contact us soon or drop by for a coffee!