Statistical vs Neural

Statistical machine translation (SMT) is done by analysing existing translations (known as bilingual text corpora) and defining rules that are the most suited to translating a particular sentence.

By feeding the SMT more data in the required languages, it will give it is higher statistical probability of outputting a more accurate translation. This also means that no human interaction is needed at any stage of the translation process. They are only required at the beginning in order to provide the text database and the calibrate the statistical models.

Neural machine translation (NMT), on the other hand, is processed through a neural network. Each neuron in the network is a mathematical function that processes data.

The initial calibration or “training” is done by feeding examples into the neural network and making adjustments based on how much error in the output there was.

This means that as the network is continually used, it will continue to fine-tune itself to provide better results.

What are the benefits?

SMT has been around longer and therefore has a wider collection of platforms and algorithms available for use. This can give it the edge on other forms of MT when it comes to accuracy of translation. Combined with the fact that less virtual space is often needed, this would mean that it is potentially a more cost effective MT system to implement and train.

However, a drawback of using SMT is that it is dependent of the quality of the source material. Bilingual text is required which may be a problem when attempting to translate less common languages.

Due to the self-learning models powering NMT, they can often be a much more reliable solution than SMT and other legacy forms of MT, especially when it comes to under-resourced languages. They are also able to better take into account context and, as a result, provide results that have a more human-like feel to them. Other advantages come in the form of speed and quality, with both increasing as they continue to learn.

Neural machine translation is also the latest advance in machine translation, meaning there is still a lot of unexplored potential. Improvements are being made all the time and being able to piggyback off advances made to artificial intelligence will be able to expedite this.

Unfortunately, like with SMTs, human input is still needed, particularly when it comes to the initial training.

With both, there will be an element of post-editing required in order to ensure that the translated outcome is fit for purpose. These normally come in two categories: light and deep.

Which one should I use?

At Prestige Network, we utilise the latest in neural machine translation to offer the fastest and most cost effective translation services.

We couple this with our professional linguists to ensure that the result is accurate and bespoke to your purpose. Our verified machine translations combine speed, cost, accuracy, and personalisation to give you and your brand peace of mind.