RUVDS together with Habr begin a project consisting of a series of interviews with interesting, in our opinion, people in the IT environment. Today we present an interview with the leader of the development of dialogue systems at Yandex, Boris Yangel.
Why Alice, not Max, what the former Minister of Communications Nikifirov was wrong about, why launch Tesla into space. Answers in video and text under the cut.
The interview was helped to prepare Lera Negrya and Habra editor Nikolay Zemlyansky.
Hey. I am Lera Negrya, and today our guest is the head of the group of the development of interactive systems in Yandex - Boris Yangel. We begin, of course, with your work, what are you doing now in Yandex?“Now in Yandex my main project is Alice, the voice assistant that Yandex launched last October, and my group is responsible for what can be called Alice’s brains. We interpret what the user said in a natural language and turn it into a structured representation. This is our first task, and the second is when we already know what needs to be answered by the user, to turn this in response also in natural language.
-
You work in this position for the past two years. All this time you have been developing. Was it a late release of a voice assistant against the background that competitors released it earlier?- No, of course, it's not too late. If we look at the competitors, who is the real competitor? There is Siri, the only voice assistant who also understands the Russian language, but it only works on iOS and MacOS, this is not the most popular platform in Russia, and there are also some questions for Siri as a product. Then, let's say, Google would have already launched its Google-assistant in Russian, would it have been worth it to get involved? Well, maybe it would still be worth it, because you can try to compete with Google on your territory for the audience, especially since we are more or less successful with the search. Well and, probably, the last point here is important: what we are seeing now is the dawn of the era of dialogue systems, and in fact neither Google, nor Apple, nor, probably, we still do not know how to do them normally. If we compare it with search engines, with what happened in the early 2000s, this is conditionally the Altavista era, when there seem to be some products, but they still work on some obscure principles and work so-so, and at any time come conditional google and change the industry completely. Therefore, of course, it is still worth going into it now. A lot of technological challenges and problems that have not yet been resolved. A new player can come and turn everything around, and we are actually in a slightly better position, because we started later, which means we have less legacy stack, that is, we can immediately work with new technologies. And for Siri, for example, which God knows when, there is a lot of old code that is already difficult to modify, and they are faced with a lot of problems because of this.
-
In your speech at the Yandex Academy you said that you caught a certain HYIP train. How did you manage it and did this train leave now?- Probably it was possible first of all due to the very bold decision that was made in our product. All competitors have some number of scenarios that are spelled out in these products: “we can answer such and such questions, talk about the weather or something else, and if what the user says does not fall under one from these scenarios, we answer with some kind of stub of the type “Sorry, I did not understand you”. ” We decided to do differently and made a neural network module that in such situations can answer something that may be appropriate in the context of the previous dialogue, and in fact it can communicate with the user on any topic. It may not always be 100% appropriate, but in more than half of the cases he can definitely say some things that need to be said at this moment. And our module has spoken a lot of such that users liked. Alice's ridiculous answers, which we did not lay there, but which she herself generated, instantly spread out on social networks, this created the very same HYIP train. Of course, he is gradually leaving, it is inevitable, all these jokes are boring, but we are trying to develop the product further and catch the HYIP, little by little, where we can.
-
You tried to make Alice as interesting as possible to the Russian user. How was the voice chosen and generally why is she Alice?- I honestly did not participate in the process of choosing a voice and I don’t know by what principles it was carried out, except that it is the cool voice of Russian actress Tatyana Shitova, which everyone loves, quite recognizable.
I can tell you about the name of Alice and about the characteristics. We thought and wrote out a number of properties that we would like to have an assistant. For example, he should not be arrogant, help a person, be friendly, be intelligent, that is, some kind of positive and negative characteristics. And then we asked people questions on our crowdsourcing platform, for example, do you associate this name with such a property or not? We looked at which names, which we chose from some principles in advance, are more correlated with positive properties than negative ones, and according to this criterion, the name Alice simply destroyed all other competitors from the short list. This is probably due to the fact that this name in the Russian culture has some positive annotations, for example, Alice in Wonderland, Alice Selezneva, classical examples. And it also concerns the phonetic properties of the name. Since this is the name that will be addressed to Alice, it is necessary that there are no sounds that will be difficult for some people to pronounce. For example, the letter "P" there definitely should not be, hissing also should not. The name Alice satisfies these properties.
“
Why Alice, not Max, why is she not a man?”- In fact, a lot of research has been done, and a woman is a safer choice. Evolution has so arranged us that a woman is a mother, something safe, dear, and a man is a warrior who just comes with a spear and slaughters you. Therefore, the female assistant is normal for both women and men, but if the assistant is a man, then this product will be less used.
-
A few years ago there was not quite a pleasant story when the chat bot Microsoft behaved completely wrong. Was there a similar risk with Alice, or did such situations even happen?- Of course it was, and we thought about this story a lot when Alice was launched. As I said before, we have a neural network module that responds in any situation, and not just within a certain scenario. And he has absolute freedom, in fact. Within the limits of some restrictions that we set, he has the freedom to speak, and he can say things in some situations that we would not want him to say. Our language is very powerful and rich, there are a billion ways to agree with something or not to agree, we will never count them all. How to prevent her from agreeing when someone is asked about something dangerous is a very difficult research task.
-
Is it somehow related to the fact that she often says: “I don’t want to talk about it”?- She says “I don’t want to talk about it” when we understand that the situation is too dangerous now for the neural network module to give an answer. In fact, if you allow Alice to say only yes or no, this is enough to offend anyone. It is enough to ask her: “Do you support X”, where X is fascism, genocide, whatever. If she says yes, write, all is lost.
-
After graduating from university, you immediately started working at Yandex; this was the department that deals with the search for pictures by text request. After that, was it suddenly a decision to work at Microsoft Research in Cambridge?
- Well, not that suddenly, it was natural. I worked in Yandex, I wanted to develop further, to do some more complex things.
-
And in Yandex did not work?- On Yandex, on the one hand, it turned out, on the other hand, I already began to lead a group in Yandex, which was responsible for the quality of search through pictures. I, probably, was not very well prepared to work as a manager, and I didn’t really like her, I loved and still love programming, and the work left very few opportunities for this, unless you want to work 16 hours a day. I didn’t like it very much, I wanted to develop exactly as a programmer, a data analyst, and then I turned up a great option, a vacancy in Microsoft, as if very relevant to my background.
-
What did you do there?- There is such a field in machine learning, Bayesian output is called. This is about how to make absolutely rational decisions in the context of incomplete information, such that one could prove that a more rational decision could not be made in this situation. I worked in a team that was involved in the creation of a software library, with which it would be possible to use all the power of this mathematical apparatus to make decisions in some situations. In fact, it is a library for machine learning, just based on some specific, non-traditional principles.
-
And the work there has prepared for a management position?- No, when I returned from Cambridge back to Yandex, I, frankly, asked not to direct anyone, said, if possible, I will write code, I really like to write code. At the same time, I was engaged in not the most important projects for the company and did not fully disclose my functionality. At first, I was developing a backend for some experimental mobile applications, then in a group of neural networks I developed text classification methods, and at some point the management decided: why would you not Boria do something useful, go do Alice.
-
Two years ago, in the press, it was actively discussed that fundamental IT education, at least in our country, is not sufficiently developed. How prepared were you? What helped you? Is this a Yandex school, or did you feel confident enough after university?- I can program because I just programmed a lot. As a child, he sat at the computer and programmed, at the university too. They all went somewhere to be drunk, but I also went to drunkenness, or I was programming and from the first year I worked in real companies almost gradually, I gained experience in the industry. This fundamental knowledge was enough to understand the other things that are needed directly to work for yourself.
The school of data analysis , of course, helped, this is a cool initiative that gives the knowledge that is not enough now at the exit from the university. I, however, spent only one year studying there, because it took a lot of time, there are so many tasks, so difficult. I was already working at Yandex, I just had no time left. But in general, I would recommend everyone to go there.
-
According to the former Minister of Communications Nikiforov, there are already too many programmers in our country, too many specialists in the IT field. Is it so? Are there a lot of really good specialists?- I don’t know how Mr. Nikiforov thinks, but as a person who needs to hire specialists directly to his team, I have enormous problems with hiring. We have a rather difficult task, and we need specialists with a higher than average level. There are practically no such specialists that I need on the market, everything that is, most likely, already works in Yandex, well, in some other number of places. I do not know whom he considers programmers, but there are very few highly skilled specialists, especially in the field of neural networks and in-depth training.
-
What do you see the meaning of your work, both in a separate company and in the industry as a whole?- Probably, if you look at things globally, it would be great to come up with something cool, a technology or an idea such that people who come to the industry after me, will develop some kind of development on this technology. That is, it will enter into the foundation on which everything gradually builds. I don’t know what it will be, but I would like to leave at least a small such mark, and it’s easier to do it, probably, if you develop as a leader, because then more resources appear, you can try some more global and ambitious ideas . But if I put a hand on my heart, I cannot say that I really like to lead people, this is really hard work, it takes a lot of time, moral strength and does not allow me to do what I like to do - to program. Therefore, I would like to develop in a direction that will allow me to find the best balance between the scale of things that I can do and which I can influence, and the ability to directly do these things with my own hands. For the industry as a whole, this is a more interesting question. It seems to me that the same thing is happening with voice interfaces now that happened with graphical interfaces, when they just started appearing, when the first color monitors appeared, and we moved from the consoles where you had to enter commands, to which some buttons press. And when they first appeared, no one knew how to do them normally, there were no paradigms, architectural patterns, frameworks, everyone did who was in that much, and the interfaces were curved, poor, it was very difficult to do them, it was impossible to read the code, but people We thought about this, and gradually some beautiful concepts emerged how to make these interfaces, convenient tools appeared, and now a professional in this area can do a website with some rather complicated interface very quickly. We think how to do the same with voice interfaces. What should be the tools so that they can be done comfortably and that they are good. Even if suddenly we did not have a good product, maybe we would have benefited the entire industry entirely by inventing something in that direction.
-
What can replace Alice?- I do not know what it will be. Which interface could be more convenient than voice ... Probably the control of the power of thought, something like that.
-
If Alice dies, will you be sad?- I will remain without work, yes, I will be sad.
-
Large companies, large projects, which are in your opinion the first candidates for departure?- Well, I probably will not call any names, it would be wrong. I can not say that someone is a candidate for departure, because he makes a bad product. Wherever I expected failures, this is where business needs any subsidies. For example, the taxi market, or some conditional food delivery - Food Fox, Deliveru Club. On the one hand, these are IT projects, on the other hand, they interact with the real world. There, in order to enter, it is now necessary to subsidize a lot, dump competitors to make very favorable prices, squeeze everyone out of the market and then raise prices. This leads to the fact that there are many who come and burn just because they run out of money. They do not have time to squeeze anyone, and mastodons like Yandex taxi in some markets of Uber crush everyone.
-
Cryptocurrencies. Invest, no, invest, what, why.- I am not the most successful cryptoinvestor, I am one of those people who decided to invest a little at the most inopportune moment, of course, and now they are sitting and hoping that all the same bitcoin will grow back. Do not do as I do.
-
Were your competitors trying to lure you?- Yes, I usually answer that now I am not looking for any other work.
-
If not IT, what then?- I do not know, I am fond of parachuting, maybe then I would become an instructor in parachuting.
-
Why launch Tesla into space?- Because it's cool.
-
Thank you very much, Boris. Our today's guest Nikolay Zemlyansky, editor Habra will ask you a few more questions.-
Hi, Boris. I want to recall the recently released film, Blade Runner 2049, where Agent K had an android assistant. I do not remember her name, let it be Masha. When they communicated and he didn’t like something, he could say: “Masha, bitchiness minus three” or “Masha sexuality plus five”. Depending on this, the communication with the assistant was adjusted. How do you think this is real in the foreseeable future and is there any point in it at all?“I’ve already talked a little about how we fight with Alice’s offensive responses.” This is all really one big story and this is why. So that Alice does not insult her interlocutor, we actually teach some models to predict whether this or that answer is offensive, and if so, to what extent. Now we use these models on a very simple principle: if the model says that the answer is offensive, we do not allow it to be answered. In the terms that you say we unscrew the offensive response to zero. In fact, we already have a model that evaluates the gradation of this offensiveness, and if there was a grocery need, we could already make such a slider that makes the answers more or less impudent. The approach we use for this can also be applied to other characteristics of the system behavior. This, of course, will require some work specific to each specific characteristic, but you can make such an assistant, whose characteristics can be adjusted with sliders, not in 20 years, but rather soon, as it seems to me.
—
, , , , , «». , . . , , ?— , , , . . . , , . , . - , , . , , , . , , , . , : , - , , . , - , .
, « » -, , , , - — - , — , . « », . .
—
, . , , , - ?— , , , , , - , - , — , , . , , . , , , , . , , - , - . — , , , .
—
, .— , , , , .
—
, , , , .— , -.
—
. ? , , , . , ?— . Google, , Google- , . , , , . , , .
—
, , ?— . . , , , . Google, , . , .
—
, Siri. ?— , -, , . , , , , , , . . , , - , . , , , . . , , , . , - . , — , -, . , , .
—
, , , , ?— , , . , , . , - — , . - , . — , , , . , Google .
—
, .
— .
