Google Translate (GT) is definitely the best and most used translation software in the world. It supports as many as 103 languages, 10,000 language pairs, and processes as many as 500 million translations every day. Experts predict that GT's neural system will soon be able to process audio and video files in addition to texts.
Google Translate since 2016, it has been using the Neural Machine Translation System (GNMT). A system based on an artificial neural network is significantly improved the translation quality. Three years have passed since then, and now we can assess its effectiveness. Has the quality of the translations really improved and what still needs to be changed in order for it to improve?
How does the Google Translate algorithm even work?
Google Translate are developed in 2006 and initially worked with help statistical methods of machine translation. That means he stored billions and billions of words in his program. When translating, he simply chose the most appropriate or popular equivalents of both languages and wrote them out. Google used in the creation of the translator United Nations documents (English, Arabic, French, Chinese, Russian and Spanish) and thus produced a corpus of six world languages with approximately 20 billion words. This process was slow, inaccurate and consumed a lot of computing power.
Today, Google Translate uses the so-called deep learning method (Deep Learning Method), in which a large artificial neural network is especially important.
Before Google started using neural networks, translation was word for word. The system easily translated each word separately, while following the basic grammar rules. Therefore, the quality of the translation was very questionable.
With the new, neural model of translation, however the basic translation unit is no longer a word, but only a part of a word. Thus, the translation is not focused on word forms, but on context and meaning of the whole sentence. The program therefore translates the sentence as a whole, according to its contextual meaning, without storing hundreds of possible translation versions in its memory.
The software therefore translates the entire sentence by taking into account the context, and no longer focuses only on individual words. It does not store hundreds of translation versions in its memory. Instead, it operates on the semantics of the text and divides sentences into dictionary segments.
How does Google translate using a neural network?
Today he uses Google Translate about 32,000 such fragments. With the help of individual decoders, it initially determines the meaning of each part of the text. It then calculates the maximum possible number of meanings and possible translations. Finally, it combines the translated segments with grammar rules. According to the developers, this approach makes it possible to ensure high translation speed and accuracy without consuming excessive computing power. However, since each language has its own rules (semantic and grammatical), Google Translate also needs special modules and dictionaries for each language, which are implemented in separate algorithms.
Interlingua
Artificial intelligence used by Google Translate as intermediate language, is called Interlingua. This universal computer language is, of course, completely unsuitable for human communication. Artificial intelligence is used in translation, where it is used as an intermediate language with which it can also translate to and from languages for which it was not created.
The advantage of the neural network is that it can operate with a larger number of language pairs, even with those that were not included in the original learning process. For example, if the system has been trained to translate the language pairs English-Japanese and English-Korean, it can also easily translate the language pair Japanese-Korean without using English as an intermediate language.
The translation method implemented by Google, the developers called zero-shot translation, is more sophisticated and relies on an intermediate artificial language for translation. This field of research is developing very quickly, and soon these translation systems will become the primary translation tool. It is a system self-taught - improves his knowledge himself and can also correctly translate slang and slang words, neologisms and other words that are not included in general dictionaries.
Language pairs
The GNMT system has greatly improved the translation of the most commonly used language pairs: Spanish-English and French-English. The correctness of translations increased to as much as 85%.
In 2017, Google conducted a survey among regular users of the translator. They were asked to evaluate three translation options: machine statistical, neural, and human. The results were impressive – the translations made by the neural network were almost perfect in some language pairs.
It is obvious that the quality of translations in the English-Spanish and French-English language pairs is almost the same as that of human translations. This fact is not surprising, since these language pairs were used for deep learning of Google Translate algorithms. The situation is different for other language pairs. However, if neural translation works comparably between structurally similar languages, then computer translation will be much worse for language pairs where the language systems are radically different.
What are the disadvantages of Google Translate?
It is true that Google Translate is very practical, and due to its accessibility and fast operation, it is helpful in everyday translation. However, it still lacks something essential – understanding. Computer translation is never focused on understanding. The developers of the program tried to improve the decryption method, or in other words, they tried to get the translation machine to handle it with its analytical abilities. However, they had to find balance between translation accuracy and speed.
Eliza effect
For any machine, computing device or software, words matter. However, machines cannot yet understand the deep meaning of words.
Years 1960 they made Eliza, a mechanical device that manipulated a series of responses, giving the impression that it was actually generating intelligent phrases. From then on, the question of whether machines can think like humans has been named Eliza effect.
The Eliza Effect has influenced AI researchers and software developers for decades. Most users of Google Translate believe that this program, at least sometimes, is able to understand the meaning of words. However, this is not true - Google Translate doesn't understand the language, but it still sometimes manages to make sentences that sound pretty good. Sometimes it even manages to translate a paragraph or two perfectly, and we almost believe that the program actually understands the language. However, we must not forget that Google Translate is not capable of thinking like a human and can only process texts in a specific way. A computer program has no memory, no imagination, and no understanding of the hidden meanings of words. So, there is no reason not to believe that computers will someday be able to think like humans.
However, they are expected to be capable of excellent translations between different languages. It is very likely that they will one day able to translate jokes, short stories, poetry and essays. After all, technology is evolving at the speed of light.
More information:
translate.google.com
