"The Rise of the Machines", "The Terminator Returns" … A lot of science fiction is built on the fact that computers are becoming so smart that they understand that they will be better off without a person. Fairy tales? How to look. By 2050, a typical home computer will be able to process as much information as all people on earth put together.
But 2050 is on condition that machines will develop gradually. It doesn't work that way. Our relationship with the electronic world is advancing in leaps and bounds. Once - a mouse was attached to the car. Once - the Internet appeared. Once - there were smartphones with tablets.
The next breakthrough is coming when the computer can understand people. Smartphones already have apps like Siri and Cortana that can have a simple conversation with us. But the problem is for the computer to understand, not what we say, but what we mean! The simplest example: the phrase “He left me”, said by a tear-stained woman and a man’s boss has completely different meanings.
So: as soon as machines learn to recognize the nuances of human speech, on the one hand, our hands will be untied in the literal and figurative sense. I talked to the computer and it did everything. On the other hand, are we not approaching a dangerous line, removing the last barrier in communication between people and soulless iron?
I have long wanted to talk about this with a well-known scientist, director of linguistic research at ABBYY, head of the departments of computational linguistics at the Russian State University for the Humanities and the Moscow Institute of Physics and Technology, Vladimir Selegey. But when I told him my apocalyptic fears, he frowned.
“Yes, science fiction is full of dark predictions about how human-trained machines will do without it. But it is not clear to me why an increase in the amount of knowledge embedded in programs will provoke a computer to make decisions without asking permission from a person?
In fact, the problem is not that the computer will learn to do without people, but that people will want to do without themselves when solving certain problems. Here is Chernobyl …
Promotional video:
And what does Chernobyl have to do with it?
- A nuclear reactor is a very complex physical model. It seems that all control parameters of the reactor are known, everything obeys strict physical laws, and decision-making can be completely entrusted to the computer. But…
In 1986, a week after the accident, I attended a seminar on the use of artificial intelligence in industry. Even then, it was clear that by entrusting decision-making to a computer, we are at serious risk. It is common for programs to contain errors. Even the satellites have software glitches.
That is, a person is more reliable than a computer, although he "thinks" much faster?
- A person has motivation in addition to everything else. He solves his problems - education, procreation, career, he knows how to feel …
But can some cunning programmer write a program that would teach an aggression machine …
- You can try to make, for example, a military robot that will make decisions by itself, analyzing what it sees and hears. And it will be very dangerous. But not because the robot will suddenly have a desire to destroy, as written by science fiction writers. But because a programmer's mistake, neglect of some factors can lead to unpredictable robot behavior.
But so far we are very far from the ability to create self-learning programs that are capable of generating completely new knowledge. Roughly speaking, independently move from the multiplication table to the ability to solve complex equations.
In the second half of the twentieth century, futuristic predictions in the field of science were very popular. Everything about the intelligence of the computer got into the milk. Nobody predicted the Internet, the incredible freedom of access to information, mobile phones. But everyone was talking about thinking computers.
For example, at the end of the 60s, our first chess system KAISA performed successfully at the World Championship among computers. It was believed that in order for a machine to beat a person, it was necessary to simulate the algorithms of a human game. To lay the mind, intellect, the very mysterious intuition that makes the chess player make the right decisions.
Today the computer beats a person. But he was never taught intuition. A billion played games, all the experience of the game, all the decisions that have ever been made by chess players were loaded into his memory. And they taught them to use this when choosing the optimal game plan, providing them with a tremendous speed of enumeration and evaluation of options. The computer beats the world champions, but does not get any pleasure from it. Everything is completely different from people.
But you are one of those who just teaches the computer to "turn on the head"
- We are only trying to teach programs to “understand” texts so that they can extract information from them, accumulate it and generalize it. So that people receive knowledge filtered, selected from billions of sources. This is very difficult, because a person himself does not know very well how his language ability is arranged, on which his understanding of other people is based.
Statistical machine translation is now popular. The computer does not understand at all what the text being translated is about, it just knows how to find the most probable variants of translating small fragments (in a few words), analyzing the huge volumes of human translations stored in its memory. In general, the text is clear. But it would be frivolous to make a responsible decision on the basis of such a translation.
Okay, can the computer translate the instructions for me?
- Instructions? Is it dangerous.
- Yes to a simple refrigerator!
- Even to the refrigerator! I would like our programs to translate, trying to understand the text, choosing between options based on knowledge, and not just because this correspondence is most often found in our database.
Okay, let's say you solved the problem and created a system thanks to which computers will learn to understand the nuances of our speech and even accurately translate it into another language. And then a new danger threatens people - there will be no need to think. There will be no need to train memory, brain. Need information - please, Wikipedia. You need to talk to a foreigner - translator …
- The higher the level of intelligence and knowledge of a person, the more useful for him the next "smart assistant". And the lower, the more opportunities appear not to think at all. Computer technologies lead to the polarization of society. Roughly speaking, the advent of the calculator did not lead to the degeneration of mathematicians. But in some of the schoolchildren, obviously, there was a decrease in the already low level of abilities.
- You know, for me, a serious indicator of a drop in the level of education is the number of people who write illiterately or are unable to express their thoughts and feelings without using mate
- Yes, people began to write less competently. Simply because, in general, there is less reading of edited texts and a lot more of such resources where spelling is very blurry. Did this make people stupider? Probably not.
Do you see a connection between the fact that a person illiterately expresses his thoughts in his native language, and the fact that he has become stupider?
“I wouldn't go that far. Although it is obvious: because children began to read less, some problems arose with the transfer of knowledge and culture between generations. This is problem. Today we see that a modern schoolchild with the same level of grades as 30 years ago knows literature worse.
Here! …
“… But on the other hand, he knows many other things much better, which no one thought they could know.
Is this a natural process?
- Yes. Moreover, for the first time in the history of mankind, knowledge can be transferred not from the elders to the younger, as has happened for centuries. There was a transfer of knowledge from the younger to the elders, which was previously not at all characteristic of human culture. A child teaches dad or mom to work on a computer, with a mobile phone, is a source of various knowledge for his parents. And even more children travel around the world. Very often they are a source of geographic and cultural knowledge.
Listen, but here it all started with computers. They freed people even from the need to memorize the elementary rules of grammar
- Just in the early 90s, I worked in a team that developed one of the first spell-checking systems for the Russian language. Was it helpful or harmful to do this? Very useful from my point of view. This system allowed for faster document creation. And then everything depends on the level of responsibility of the person. Who said that after the car there is no need to do a check at all? The system only helps to make it much more efficient.
But we trust the computer
- It only means that the person who made the program did not inform you that there are many phenomena that the machine cannot verify. Negotiation, for example.
New technologies, alas, often lead to the loss of traditional skills. I've been doing repairs in the apartment. I wanted to install wooden windows, which were installed before. But this turned out to be impossible. All are installing double-glazed windows. People, alas, stop doing many things with their hands. What to do, that's how life works. Unfortunately.
So you are a conformist?
- Not. I believe that doing absolutely harmful things is not good. But when you have both gain and loss, you need to assess the risks … If you make drugs, you do not think that not only good people, but also bad people will survive as a result of its use.
In our case, simultaneously with the development of technologies, it is necessary to engage in training of those who use them. The negative consequences of new technologies are reflected primarily on those who do not do their job very well at any technological level.
Take medicine, for example. We want to help the doctor make a decision, we make a computer expert system based on the analysis of a large number of reliable diagnoses made by the best doctors. It gives a professional the opportunity to turn to a larger volume of knowledge than he himself has. But the final decision is his, not the computer! For a bad doctor, everything is different - he will trust what he cannot reliably check. But on average, with the advent of such technologies, it seems to me that medicine is still getting better, not worse.
Well, we were convinced that there will be more benefit than harm from the intellectual system of understanding a person by a computer. But when will she appear?
- This is a very difficult task that cannot be solved immediately. It is necessary to teach the computer knowledge about language and knowledge about the world, methods of inference and comparison of values. Something we know how to do now, something will take years. For example, we need reliable data on language use, taking into account individual and social differences. And these differences are very significant. For example, we made a special project dedicated to regional differences in the Russian language. Collected a dictionary, in which almost 10 thousand words, which are the norm for residents of only certain regions of our country
Who are "we?
“This is a joint project called Languages of Russian Cities, involving ABBYY specialists, linguistic scientists, enthusiasts from different parts of Russia and Russian-speaking countries. We analyzed the language of regional media, social networks, decrees, and decisions of local authorities. And this is only a small part of the knowledge that needs to be taught to a computer!
What is the fundamental difference between your approach to translation and the systems that Google uses, for example?
- Statistical systems do not build linguistic structures. They look for translation matches for small fragments of 5-7 words. But the language is so arranged that very often the related words are located much further, and if this connection is not taken into account, errors arise. But very often such details are important in translation, neglect of which can completely change the meaning. In order to take everything into account, it is necessary to identify during the analysis the entire system of linguistic connections between words, what is called the structure of a sentence. We do not translate pieces of sentences from language into language, we are trying to identify the semantic structure of the sentence, and then synthesize this structure by means of another language.
But these semantic structures are used not only and not so much for the purposes of machine translation. They are also needed to effectively solve the problems of intelligent search and information analysis, which are more important today for ABBYY as a commercial company. For example, the recently released first solutions for the corporate market are based on these technologies: they just allow you to search and analyze data in the huge flow of information that is stored in organizations.
And which language, from your point of view, is more difficult to understand?
- It is generally accepted that the more difficult are those languages in which you need to learn more rules. But in the case of computer understanding, this is not the case.
Something that is difficult for a person, for example, a rich inflection system in Russian or Lithuanian, simplifies the task for a computer at certain stages of analysis.
For example, the Chinese language is very difficult because there is no morphology in it, and hence the large amount of ambiguity that the computer faces. Therefore, for computer analysis of the Chinese language, it is very important to use knowledge about the world, more important than, for example, when analyzing languages with a developed morphology.
We managed to work with Russian, German, English, French, Spanish, Chinese. Language analysis is divided into stages. Some stages have specificity, difficulties for one language, some for another. But in essence, the technologies we work with are applicable to all languages.
In your opinion, when will the day come when we will be able to receive a quick and understandable translation on the computer in parallel with our conversation. Let's say we're talking now, and immediately get a transcript of our conversation in Spanish?
- It will be possible pretty soon. I don’t think that in two years. But … What is needed to solve this problem? We need to improve our speech analyzing systems. We cannot yet take a spontaneous conversation, separate it from the noise and get the desired text. And there is progress in this direction every year.
But all the same there will be such things that the machine cannot interpret unambiguously. And here the accuracy of understanding and translation will depend on the availability of the system of knowledge about the world, on the mechanisms of logical inference.
But in any case, the danger of complete trust in such a program will remain. It is a mistake to think that, say, in five years you can dictate a letter to a computer, it will translate it, send it to your business partner, and everything will be fine. No one guarantees that the machine will not miss the "not" somewhere or make other mistakes, after which your business relationship can be considered complete.
That is, a person as a controller should still be?
- Certainly. Trusting technology blindly is dangerous.