It is worth noting that lately, not only machine translation technologies, but also artificial intelligence in general, have made a huge step forward: you will no longer surprise anyone with the phrase “neural network” or “self-learning system”.
However, this was not always the case – for a long time, scientists tried to create a mechanism for translating from one natural (this is an important condition in the definition of machine translation) language to another, but the attempts were unsuccessful.
Types of machine translation
Machine translation systems fall into three categories: grammar-based systems, statistical systems, and hybrid systems that combine the advantages of the first two groups.
There is another, relatively new type of machine translation – neural. It will be discussed in more detail later, but for now let’s look at the first two types.
The text is analyzed and its translation is built on the basis of built-in dictionaries and a set of grammar rules for a given language pair. PROMT and Systran are the best known examples of RBMT systems.
The quality of such translations leaves much to be desired, but they are still used.
Among the advantages of RBMT are morphological accuracy (words are not confused during translation), reproducibility (always the same result) and the ability to customize the system for the subject area (teach special terms).
The principle of statistical analysis is applied: huge volumes of texts in the source language and their translations made by a person are loaded into the program.
The program analyzes the statistics of interlanguage correspondences, syntactic constructions, etc., then relies on it when choosing translation options – this is self-learning.
This is where neural machine translation comes into play. self-learning is typical, first of all, for translation based on neural networks. This type of translation began to appear in the 1990s and is now the main type of machine translation.
What is a neural network
A neural network is a mathematical model built on the principle of networks of nerve cells of a living organism. The ability to learn is one of the main advantages of neural networks over traditional translation algorithms.
The system can also be trained by a person, correcting the translation results. This is exactly how Yandex and Google online translators work. Thanks to self-learning, the quality of their translation improves with each new translated text.
Over the past few years, neural networks have surpassed everything that has been invented in translation over the past 20 years. They even learned to agree on genders and cases in different languages (themselves!).
In addition, for the first time, it became possible to directly translate between languages that did not have a single common vocabulary. Previously, statistical translation methods always worked through English. Neural translation does not need this.
The hybrid method has a number of advantages. For example, neural translation does not always do well with short phrases. A simple statistical translation is usually better at finding the equivalents of set phrases.