David Yang (ABBYY) on how FineReader was created
I was recently at a meeting of a certain club, where David Yang spoke (founder of ABBYY). Part of the speech recorded on the phone. Decryption records present to your attention.
If a company enters the market and there is already something similar in this market, then you can read about it in many books, in particular, on the conduct of military strategies. The basic idea is as follows: In order not to face a strong competitor that is called "in his field." That is, you need to find your battlefield, where the competitor is completely defenseless and was never going to even expose the forces.
I will give an example: at one time we were leaving with a product called Fine Reader. In 92, Kostya Anisimovich, our technology officer, approached me and said, "let's make our own recognition system." It looked absolutely like madness, because we ourselves sold another recognition system called “Authors”, we were its dealers.
When selling this system, we faced the fierce competition in the Russian market, there were programs called Tiger, Uniform, Intuition, and some other Russian-made programs. There were programs of Western production (lists), I can name a dozen strong Western systems, which then existed and were leaders in the international market. And the words “let's make our own” looked like madness, because it meant: let's invest super big money, a lot of resources, our best talented heads with a completely unknown result.
')
The question is, how are we going to be better? What can we do that which they have not done so far? Because we are inferior, they have a head start over several years of development, multiplied by dozens of people in the development team, that is, they have several dozen man-years. And of course it is very scary to engage in such a struggle.
But since we came out of the Physics and Technology Institute and we always felt a knee-deep sea, we somehow scratched our heads and said, "Ah, we will."
We actually began to act as follows. We parallel planted two teams. One team developed Finereader 1.0, another team developed Finereader 2.0. Absolutely two different teams. The first team had a task to quickly make a recognition system, and to get out, it was a very important moment, what is called time to market, that is, it was necessary to reach these positions quickly. What are the positions?
We found that, there was no need to go far, the main parameter by which the recognition systems were evaluated was the recognition accuracy. That is, how much the program makes mistakes when entering text. And on this most important parameter, we had to lose, because it really requires a man-fly on artificial intelligence technologies associated with the recognition of various letter shapes, printing defects, paper jams, poor photocopy and all that. We most likely should have lost on the most important parameter, how can we fight?
We found out that in fact people in Russia are scanning technical texts, an economic plan, bilingual texts. That is, along with the Russian main text, Latin terms, brands, etc. are sure to be found. And the existing systems, yes they were then leaders in the accuracy of recognition of a monolingual text, as a result they replaced all the words of a foreign language, roughly speaking, with asterisks, they had to be interrupted again. We decided to make an intelligent bilingual system that will automatically distinguish the Russian letter y from the Latin letter u by context, or the Russian letter p from the Latin letter p which cannot be visually distinguished, only it can be distinguished by context.
We will make an intelligent system that will recognize bilingual texts. Yes, it will make more mistakes, and it made more mistakes at the beginning, but integrally, when the bilingual text was introduced, it actually contained fewer errors in this understanding, fewer asterisks and it turned out to be very popular in the market. The second thing that we immediately put into the system is the so-called unishrift, that is, it did not need to be taught. Then the old recognition systems had to be trained in the font before proceeding to recognition. There are still some elements, she was the first on Windows.
It was absolutely risky to go out with this proposal after a year, and after two years, because our competitors breathed into the back of their heads, and they had to release the same bilingual, the same for Windows, the same uniphonte system there after two years. We could only go out in a year. We found methods, together we cooperated with a group of developers. And they did it literally in 10 months. It was hard to believe, but after ten months, in August 1993, we sold the first 40 copies of Finereader 1.0 which had these functions.
Thus, taking a position in the market as the first in the market under Windows, intellectual, without tuning to the font and bilingual.
And began a fierce struggle, tests in the press. The competitors were absolutely stunned by the fact that we were out of the corner. It would seem that we have less development experience, and so on, but we found that clearing, that market niche where they turned out to be completely unprepared. All these years they improved their accuracy of recognition in one language. And while we fought with them in the press, while the press then compared the monolingual texts we lost, it compared the real bilingual texts we won, while all the battles took place there, another group was involved in Finereader 2.0
And after three years, we finally released Finereader 3.0 which was completely new, which had all the advantages of Finereader 1.0, but at the same time it was already very competitive in its main parameter, the recognition accuracy of a monolingual text. And at that moment we suddenly discovered that it is better not only in Russian texts, but it is better than the American systems in English texts. A year later, we started selling it abroad under a different name. ...
At some point, we began to win in all respects. The last parameter, recognition speed, we limped over this parameter for some time. Because yes, we had higher quality, but we lost in speed. But the speed of half a minute or one minute of recognition, when the typist spends 15 minutes on the same page, it was actually not so significant for the user. As a result, we are absolutely at the edge of the abyss, having traveled these several years with a large group of developers, until we have established a solid and in all respects. By the recognition accuracy, by the number of languages, by the convenience of the interface, by almost all parameters, we began to be ahead. Well, we overtook our Russian colleagues in the late 90s, and in the early 2000s we took a strong position in the European and American markets. And effectively, now in terms of the number of licenses that are supplied in the world, we are leading the recognition systems ...
This is the story of how we managed to find this glade at that moment.
P.S. Continuation will follow if it turns out to be interesting to anyone. Just do not want to decipher, the process is quite time consuming.
--------------------------------If a company enters the market and there is already something similar in this market, then you can read about it in many books, in particular, on the conduct of military strategies. The basic idea is as follows: In order not to face a strong competitor that is called "in his field." That is, you need to find your battlefield, where the competitor is completely defenseless and was never going to even expose the forces.
I will give an example: at one time we were leaving with a product called Fine Reader. In 92, Kostya Anisimovich, our technology officer, approached me and said, "let's make our own recognition system." It looked absolutely like madness, because we ourselves sold another recognition system called “Authors”, we were its dealers.
When selling this system, we faced the fierce competition in the Russian market, there were programs called Tiger, Uniform, Intuition, and some other Russian-made programs. There were programs of Western production (lists), I can name a dozen strong Western systems, which then existed and were leaders in the international market. And the words “let's make our own” looked like madness, because it meant: let's invest super big money, a lot of resources, our best talented heads with a completely unknown result.
')
The question is, how are we going to be better? What can we do that which they have not done so far? Because we are inferior, they have a head start over several years of development, multiplied by dozens of people in the development team, that is, they have several dozen man-years. And of course it is very scary to engage in such a struggle.
But since we came out of the Physics and Technology Institute and we always felt a knee-deep sea, we somehow scratched our heads and said, "Ah, we will."
We actually began to act as follows. We parallel planted two teams. One team developed Finereader 1.0, another team developed Finereader 2.0. Absolutely two different teams. The first team had a task to quickly make a recognition system, and to get out, it was a very important moment, what is called time to market, that is, it was necessary to reach these positions quickly. What are the positions?
We found that, there was no need to go far, the main parameter by which the recognition systems were evaluated was the recognition accuracy. That is, how much the program makes mistakes when entering text. And on this most important parameter, we had to lose, because it really requires a man-fly on artificial intelligence technologies associated with the recognition of various letter shapes, printing defects, paper jams, poor photocopy and all that. We most likely should have lost on the most important parameter, how can we fight?
We found out that in fact people in Russia are scanning technical texts, an economic plan, bilingual texts. That is, along with the Russian main text, Latin terms, brands, etc. are sure to be found. And the existing systems, yes they were then leaders in the accuracy of recognition of a monolingual text, as a result they replaced all the words of a foreign language, roughly speaking, with asterisks, they had to be interrupted again. We decided to make an intelligent bilingual system that will automatically distinguish the Russian letter y from the Latin letter u by context, or the Russian letter p from the Latin letter p which cannot be visually distinguished, only it can be distinguished by context.
We will make an intelligent system that will recognize bilingual texts. Yes, it will make more mistakes, and it made more mistakes at the beginning, but integrally, when the bilingual text was introduced, it actually contained fewer errors in this understanding, fewer asterisks and it turned out to be very popular in the market. The second thing that we immediately put into the system is the so-called unishrift, that is, it did not need to be taught. Then the old recognition systems had to be trained in the font before proceeding to recognition. There are still some elements, she was the first on Windows.
It was absolutely risky to go out with this proposal after a year, and after two years, because our competitors breathed into the back of their heads, and they had to release the same bilingual, the same for Windows, the same uniphonte system there after two years. We could only go out in a year. We found methods, together we cooperated with a group of developers. And they did it literally in 10 months. It was hard to believe, but after ten months, in August 1993, we sold the first 40 copies of Finereader 1.0 which had these functions.
Thus, taking a position in the market as the first in the market under Windows, intellectual, without tuning to the font and bilingual.
And began a fierce struggle, tests in the press. The competitors were absolutely stunned by the fact that we were out of the corner. It would seem that we have less development experience, and so on, but we found that clearing, that market niche where they turned out to be completely unprepared. All these years they improved their accuracy of recognition in one language. And while we fought with them in the press, while the press then compared the monolingual texts we lost, it compared the real bilingual texts we won, while all the battles took place there, another group was involved in Finereader 2.0
And after three years, we finally released Finereader 3.0 which was completely new, which had all the advantages of Finereader 1.0, but at the same time it was already very competitive in its main parameter, the recognition accuracy of a monolingual text. And at that moment we suddenly discovered that it is better not only in Russian texts, but it is better than the American systems in English texts. A year later, we started selling it abroad under a different name. ...
At some point, we began to win in all respects. The last parameter, recognition speed, we limped over this parameter for some time. Because yes, we had higher quality, but we lost in speed. But the speed of half a minute or one minute of recognition, when the typist spends 15 minutes on the same page, it was actually not so significant for the user. As a result, we are absolutely at the edge of the abyss, having traveled these several years with a large group of developers, until we have established a solid and in all respects. By the recognition accuracy, by the number of languages, by the convenience of the interface, by almost all parameters, we began to be ahead. Well, we overtook our Russian colleagues in the late 90s, and in the early 2000s we took a strong position in the European and American markets. And effectively, now in terms of the number of licenses that are supplied in the world, we are leading the recognition systems ...
This is the story of how we managed to find this glade at that moment.
--------------------------------P.S. Continuation will follow if it turns out to be interesting to anyone. Just do not want to decipher, the process is quite time consuming.
Source: https://habr.com/ru/post/88852/
All Articles