Of the total languages Google Translate can translate by typing, there are 103 languages. In September 2016, Google announced that it is developing translations using Neural Machine Translation (NMT) to help translate languages better. Have been using this system since March 30th.
From the beginning, Google Translate translates sentences or
texts in Thai language to understand some. do not understand Sort of some
unknown words, as Google invented machine translation that uses massive
statistical models to identify hundreds of millions of data content patterns on
the web. When Google Translate launched Google Neutral Machine Translation
(GNMT) in 16 pairs, Languages with English such as Thai, Chinese, Japanese,
Korean, Russian, Hindi, French, German, Spanish, Portuguese, Turkish,
Vietnamese, this tool helps translate and communicate in a more natural way. It
will translate the whole sentence at once. instead of translating part by part
This will use a wider context to help translate it to the most relevant
meaning. The system then arranges and adjusts to match or closest to human
spoken language. Multiple text translations Paragraphs or articles are easier
to read and understand.
It
can be done because people teach the machine first.
Michael Jittiwanich, Head of Marketing, Google Thailand, said that Google Translate translates better because it uses a neural system developed from machine learning technology, a method that works within the framework of AI (Artificial Intelligence). one is smarter As Michael explains, Machine Learning is about making machines learn, improve and make themselves smarter. For example, Google does machine learning because it wants to make the system understand 'What people are asking' and 'what the user is looking for'. For example, Google sorts the information they have in the world, so users can access and benefit from the information around the world. Access to information in many ways, such as typing a message sent into the system. speak to make a sound Take a photo and send it in This allows the system to better understand what people are doing or wanting.
The principle of machine learning is to learn from repetition.
In the past, computer programming would write logic step-by-step. For example,
if this happened. The computer must do this. But there is a limitation that
what computers can do There is only what the programmer wrote. Programmers
can't fully predict what will happen, so commands can't cover all potential
steps.
But in the age of machine learning, programmers teach
computers how to learn on their own from the experience of repeating learned
systems. Computers will understand that and find patterns in their own work.
without the programmer having to tell every step.
An example of learning is when you see pictures of dogs and
cats. If it is a human brain, it will immediately process whether it is a dog
or a cat. But the computer will judge from the picture. see color information
by pixel It may not be immediately identified as a dog or a cat. Therefore, the
code was written to give computers a system that mimics the thoughts of the
human brain for better processing. The human brain has a network of neurons
called Neurons that respond to touch, sound, and light. Computers are designed
to have a neuron connection style. The computer will see a pattern that can
capture the characteristics. And it is a special ability of the system that the
old coding will not see.
When a computer has a small computing unit as a neuron, the
connection of a neuron forms a neural network which sees more layers, and when
there is a neural network, it has to teach with a large amount of data. The
training method is leading. Millions of images of dogs and cats are sent into
the Neural Network one by one, and the computer learns and thinks. What is
caught? The first layer receives the information and then passes it to the
second layer and the third layer is recognized step by step as the body, eyes,
ears of the cat. Then the system will find patterns that further north to say more
clearly that Which picture is a dog? Which picture is a cat?
“In order to tell whether each figure is a dog or a cat, A
system that makes a very good guess at a certain level has to be mastered
hundreds of millions of times. This concept has been around for a long time.
But it has only been available in the past few years because computing
technology of computing systems in the last five to ten years is much faster
than before. So large amounts of data can be processed in no time,” said
Michael.
And the ability that is superior is that when the layer is
made higher and higher Computers can see and understand abstract patterns, such
as images of hugs. regret pictures The system will understand abstract concepts
by looking at the pattern's layers continuously, which mimics the activity of
the brain that people who see it repeatedly will remember it.
This neural network learning trick is used by Google to
improve Google Translate's translation skills by having to input a huge amount
of linguistic information in order to better recognize and understand the
language.
Brings
out the uniqueness of matching more than two languages.
When teaching a machine to understand one language, Google
Translate requires at least a hundred million text and language data from Google
to run the data to train in pairs with another language. Teaching one language
takes about a few weeks. If 103 languages are paired one at a time, Google takes on epic scale data. That
would take a lot of time and resources. Google trained Machine to translate
language pairs in Zero-Shot Translation by inventing a new model. 'Multilingual
Model' is to combine languages that are close to each other and translate
them into layers at the same time by doing multiple languages at a time, such
as translating Japanese > English, translating English > Korean, which
when Machine understands Japanese well. climb Korean language will also be
better because it is in the same model, resulting in higher quality data and
when testing machine to translate Japanese into Korean directly. I found that
the translation was good enough. although there was no direct translation data
between the two languages before.
Barak added that at first Google Translate tried to match Thai
language with Vietnamese language, but with the data, ASEAN languages have
less Traning Data that the translator sends into the system to train the
machine than the European side, so try to use Polish language, which is used by
many people. Very well developed and used to match with Thai language, found
that Neural Machine Translation works and learns better.
“These two languages are not very close to each other, but
Machine finds that the origin of the language is High Level Similarity.”
Use
behaviors and gaps that need further development
Google Translate is available through 4 platforms:
translate.google.com, Chrome Browser, Android and iOS applications, and
Enterprise API. Making games, creating applications.
Barak Turovsky (Barak Turovsky), Head of Product and Design,
Google Translate Google, Inc. provides information about the behavior of using
Google Translate:
- Every month, Google Translate is used by around
500 million people worldwide.
- Data is translated 1 billion times a day, totaling
more than 140 billion words a day.
- About 50% of the web is in English, while 20% of
the world's population speaks English. It is therefore not surprising that 95%
of all translations occur outside the United States. especially in developing
countries.
In terms of translation quality, Google had linguists
experiment with three different types of Google Translate quality data for
certain languages: Human-Translated, New Neural Translation (GNMT), and Old
Phrase-based Translation (PBMT). The quality grade was 0-6, it was found that
the good translation of people was not yet at the level of Perfect Translation,
such as Chinese-English translation. The quality of the human-translated
language was almost a grade of 5, while the neural translation (GNMT) was
slightly graded 4. It reflected that the quality was close to that of a human-translated
language. while translating with Phrase-based system got a 3rd grade,
reflecting the translation improvement that Google Translate has done better
than before.
“Neural Translate has evolved by leaps and bounds in the past
10 years, even if it is a whole sentence. But before, it was translated word by
word with the same system. Now it's more like the human brain is translate
whole sentences at once By re-arranging the sentences to see if it makes sense
or not in order to get the most understanding. In the meantime, the machine
will learn all the time. For example, Thai language should be arranged and
flowed. How is Grammar? It will be close to what people do. talk more and more
But of course there are gaps that need to be developed, which will see
improvements closer to the language people use,” Barak said.
The reason why human translators have not reached Perfect
Translation, Google translate explains that because the translation depends on
the Context, the choice of each person's vocabulary. Because knowledge and
preferences in translation are not the same. And even the same person's
speaking and writing style are not the same. So it's difficult to make translations
Perfect for everyone.
Google
Translate next step
Google Translate is also aiming to improve the translation of
short sentences, Brand, Idiom, year into pounds to kilograms. kilometers to
miles, etc. to be better. It is also open to the general public to edit or add
information to the Google Translate system via Suggest an edit, which is a
Cloud Source Platform and Machine can learn languages through this.
In terms of using the offline translation application, Google
Translate app can currently be translated into 52 languages, requiring users to
download a database of about 25 MB in one language into their device first. The
Neural Machine Translation system is also does not apply to the application But
it is developing on the base of 25 MB Database.
Google uses machine learning in every product to make some services better, such as Search Results, Real-time Translation like understanding images from a mobile phone camera and translating the text seen, Google Photos like searching for the word Beach. You'll find pictures of the beach, and Google is developing machine learning to understand videos that are more complex than images.
“The technology trend that will come in the future will be
machine learning to make a lot of good services, some of which we will see
clearly, such as image search from machine learning, but many of them are what
we are using. already at present but got better results,” Michael concluded.
read more: What is the most effective treatment for Labyrinthitis?