Answers of the general director of ABBYY Sergey Andreev to questions of readers of Habr. Part 1

Dear readers and writers of Habr!

I want to say thanks to all participants in this discussion for their sincere desire to make our country better. This can be seen in all the questions asked here, even if they sometimes sound harsh and prejudiced. I considered it important to answer the questions, because ABBYY had not talked about what we do, or how we do it for too long. And the lack of information gives rise to a lot of ridiculous guesses. So, time to answer questions.

In these answers I will sometimes be distracted by general judgments and general information, but not in order to advertise something, but in order to better explain the reasons underlying our actions.

To begin with, ABBYY is already 21 years old. All this time we have been making interesting products and technologies that are known all over the world. Now these are printed and handwriting recognition products and dictionaries. In the future, these are products based on computer linguistics. Our interests do not lie in the field of consumption. Our leaders and shareholders do not drive luxury cars, do not buy yachts, helicopters and palaces, we do not have summer cottages on Rublevka. We are interested in making amazing new products, rather than going shopping.

More than 70% of our income is not earned in Russia. According to various estimates, from 30% to 50% of scanners and MFPs all over the world are equipped with the Russian software product ABBYY FineReader. We are pretty well diversified by source of income. We have no dependence on the Russian market and, especially, on its public sector.

We do only what interests us. Only by what we are ready to spend time on our lives. We are not interested in cuts, politics, corruption schemes and other muddy activities.

We live in Russia, work in Russia, pay all taxes. Somehow they even received a letter from the tax inspectorate as the best taxpayer in the area :) (I don’t know whether to be happy or worried). Our children study here, and nobody is going to leave anywhere. Everything we have is here! We do not care what will happen to Russia. And we are ready to participate in useful affairs for the country. We spend our time and money on the development of education and on other useful initiatives that are developing the IT industry in Russia.

It may all seem to someone too exalted, but here we really have a reserve of good people with good and common interests.

Now let's switch to the substantive questions. There were many questions, many answers, so today only the first block will appear here so far. In it, I will answer questions about ABBYY Compreno technology, for which ABBYY received a grant. The following blocks will be published early next week.

"... Money is allocated to create Compreno automatic word processing technology ... Compreno technology is designed to create systems for analyzing, translating and searching texts in various languages. Using this technology, for example, it will be possible to automatically compose a context database for improved text analysis, and auto-abstract texts and etc..."

I would like to hear more konretika, with explanations on examples. I guess that Compreno technology can be used in computer translation of texts, for example, to correctly determine the structure of sentences in the source language - that is, to determine the relationships and relationships between words in a sentence in the source language. Thus, the generation of sentences in the final language will be carried out taking into account the entire structure of the original sentence. As far as I can see, in modern auto-translators such an analysis is in its infancy, approximately at the level of definition of a noun (pronoun) and a verb to it. Therefore, the proposals of modern translators are quite clumsy.

How much will the quality of the translation improve? Is it possible that sentences will be generated linguistically related, with words with correct endings, that is, at the level of structure - there will be no complaints. But at the level of meaning, how did the auto-translator generate nonsense, and will continue to generate pseudo-meaningful text?

Answering the first two questions, I would like to tell you more about ABBYY Compreno technology.

Although we are closely following developments in this area in the world, we are not aware of the analogues of the new generation of linguistic technology ABBYY Compreno that we created.

The central core of the technology being created is a universal hierarchy of concepts and a model of relations between these concepts (for specialists: a hierarchy of universal semantic meanings and relations between them). Although all people on earth speak different words, they use a very similar system of concepts. People in different countries go to work, sit at home, work on computers, draw up contracts, fly on airplanes, and negotiate. Similar business centers are being built for them. They are located in similar rooms and use similar furniture. All these concepts and their relationships in different developed civilizations have much more in common than different. In the future, I will call this semantic concept tree the English abbreviation USH (Universal Sematic Hierarchy).

USH is a concept tree that is universal for all languages, the thick branches of which are more general and universal concepts (for example, "travel"), and the thin ones are more specific, but also universal concepts (for example, "business trip"). The wood structure makes it possible to ensure the inheritance of properties from ancestors to descendants, thanks to which the description of new concepts is faster, since in order to describe the concept of “order”, you no longer need to list all the characteristics of the concept of “document”. Words of a particular language are leaflets on the USH tree. We get the opportunity to disambiguate. For example, in Russian, different meanings of the word “management” correspond to concepts on different branches of this tree, because there is a meaning of “management” as a department, and there is a meaning of “management” as an action.

The semantic description of a particular language turns into the attachment of “leaflets” - words of that language - to USH branches.

The second, but no less important part of the technology is the complete parsing of the text. Syntax is a way of “coding” meaning (for specialists: semantic relations) in a particular language. The semantic relations themselves are universal, and the ways of their implementation in each language are their own. In some languages, a linear order is established, in others cases, prepositions, special service words are used, somewhere all at once is used. For each language, the syntactic description is done anew, but the means themselves, which different languages ​​use to encode the meaning, are enumerable. When describing a new language, different constructor elements are used (the same linear order, different types of syntactic transformations, grammatical meanings, prepositions, special constructions).

Compreno technology also successfully defines more complex syntactic relationships, such as replacing the word “boy” with the word “he” in the sentence (for specialists: anaphora): “Although the boy wanted to play, he understood that he had little time.” Or whole gaps in complex sentences (for specialists: ellipsis), for example, "he loves red wine, and she likes white wine." The connections between the concepts distinguished by the system are also expressed in a tree structure, they actually convey the meaning of the written, and carry important information for search or translation. Thus, the system seeks to determine the meaning of a text written in an ordinary language, allowing the machine to "understand" this text and transform it into a universal representation, independent of the language.

Using USH, a syntactic description of the language, as well as statistics of the relationship between words, Compreno technology performs a complete analysis of the text and, when translating it into another language, uses the words corresponding to the correct branches of the USH tree and the relationships identified when parsing the original sentence.

As you understand, if for translation we managed to bring the computer closer to understanding the meaning of the text, then this understanding can be used not only for translation tasks, but also for many other extremely popular applications. Obtaining a universal representation (meaning) makes it possible to come close to better speech recognition, a smart search for information when using natural language as a query, and as a response you can get a document that does not necessarily contain query words, but contains their analogs and correct ones relations between concepts. You can determine the authorship of a document, you can make an abstract (squeeze from a large document). There is much more to do when you have ABBYY Compreno, the universal basic linguistic technology.

What problems does the product for which the money is allocated in Skolkovo solve? What are the areas of its application? What new opportunities will he provide to ordinary people?

List specifically, in what types of programs do you now see the need to use Compreno, if suddenly now it would be completely ready?

Again I will answer two questions at once.

Compreno technology is a universal linguistic platform for applications that solve many applications of natural language text processing, such as:

1. Translation and interpretation from one language to another;
2. Intelligent search , in particular:
   
   • Search by meaning, not by keywords;
   • Extract facts and relationships between search objects (including for competitive intelligence);
   • Monitoring companies and personalities and building analytical reports based on various types of parameters. For example, when preparing a report on the topic “What is the most popular mobile operator tariff?” It is important not only to correctly identify all the tariffs discussed in the media, but also to compare the frequency.
   • The ability to receive answers to queries asked in the usual language (for example, “What does Ivan Ivanovich Bobrov own?”).
3. Multilingual search i.e. when a question in one language is answered in all languages ​​supported by the system;
4. Classification and filtering of documents;
5. Protection against unauthorized use of information;
6. Automatic abstracting and annotation of documents;
7. Speech recognition;

Nowadays, there is no such universal technology in the world that allows solving so many applied problems requiring high-quality linguistic analysis of texts. And we are talking not only about traditional tasks, such as translation, but also tasks whose solution at a qualitatively new level was impossible in the past (for example, the automatic search for facts and relationships in information arrays).

The revolutionary nature of ABBYY Compreno lies in the fundamental approach. Many people thought of a universal system of concepts and technologies of full syntactic and semantic analysis. In our work, we rely on the works of leading Russian scientists in this field and classical linguistic education. However, many experts retreated to the colossal engineering and linguistic complexity of implementing this idea for real practical problems. Advanced Russian linguistic education and science have given us a very good foundation for the beginning and development of this great work.

What will the money be used for? What are the expected results? Will the result be a new product, or will it go old?

The project has been under development for about 10 years. I don’t know about any public results. Has he turned into a long-term construction with unclear prospects? What are the guarantees that this grant will be the last necessary for the release of the product?

How long is the project to be implemented in Skolkovo?

I answer three questions.

In psychology, it is known that a child is distinguished from an adult by an inability to control his impulses - and this is one of the reasons why he can never replace an adult in many types of adult activity. A mature company built for a long time is also distinguished from a one-day company by its ability to invest in projects that will bear fruit over the years. Read more here .

I will add that we have long decided that you need to do only what you are sure that you can achieve the best results in the world. If you have no reasonable reason to believe that you will become the best in the business in the world, then this is bad business, because it will ultimately be reduced to price competition. Unique things are not quickly done, otherwise they would be easy to repeat.

Well, now, in essence, what we are doing.

Work on the creation of Compreno technology has been going on for 15 years (you can look at the people who are involved in this and our other projects here) The fact is that this project requires a serious scientific foundation, without which it is impossible to create a working technology of really high quality. Fundamental science, as you know, requires money and time. 15 years ago, thinking over the basic concepts of the new technology began. About 10 years ago - work on building the architecture, about 6 years ago - serious work on programming the basic modules, and about 2 years ago the most serious technological risks were passed. The project has entered a phase that allows us to speak with a high degree of confidence about the attainability of our goals.

The result we are striving for is described in one of the previous answers.

We expect that commercial products for widespread use on the basis of the Compreno platform will appear within 2-3 years. But today we are demonstrating Compreno technology to large customers. Experts who see how the technology works today are discussing pilot projects with us. In addition, we are preparing another product of ours - a library of functions available for other application developers. That is, any developer will be able to license the core of the system and integrate these functions into their software products.

In the end, in order to remove doubts about the long-term construction prospects, I want to add that all these years, the company's shareholders (roughly coinciding with its management), instead of enriching and buying yachts, villas and other rubbish, have invested the company's profit in a completely new breakthrough direction. We were able to build a high-quality science-intensive international business, and we consider ourselves competent in what we do. Who better than us to know where to invest our money? Can anyone other than us more reliably evaluate the correctness of this investment? If someone believes that we are wrong, and knows how to use our money better, then I can congratulate all of us on the emergence of new businessmen who will glorify our homeland with their deeds. Dear young! We are waiting for posts about your success!

I would like to know more about the recognition system, namely. Will it be "language independent" or will it also be based on the morphology of each language, like the current system in FineReader? In the second case, is it planned to use an open format or technology that allows users to add morphology rules for new languages? Say, for example, the hunspell / aspell dictionaries in browsers.

I hope the previous answers have clarified this topic. If not, please specify.

What is the profit to the state and citizens from this technology? I am interested in its use in something other than your products.

Few know that selling ABBYY FineReader feature libraries for developers brings us more revenue than selling ABBYY FineReader boxes. We also plan to make ABBYY Compreno technology available to other developers so that it can be included in other products that require high-quality text processing in a natural language. This will spur the development of a wide variety of Russian businesses using Compreno to create programs or provide services. We believe that part of these high-tech products or services will be sold abroad, which in the current situation of predominantly Russian oil and gas exports cannot but rejoice.

In addition, we must not forget that the creation of such a technology here in Russia increases the general level of qualification in this area in the country. Because people who understand this so well in applied linguistics live, work, pay taxes and spend their money in Russia, and not in California, Munich, Calcutta or Guangzhou. Conferences, lectures and seminars held by ABBYY on the topic of artificial intelligence and applied linguistics are held in Russia. ABBYY supports education in this area at our universities, and not at MIT or Peking University. In general, Russia has a chance to gain world leadership in this local area of ​​knowledge, and not America, Israel or India.

Immediate PROFIT from all this is difficult for the state and citizens to calculate, but here the BENEFIT is very large, especially if you look at the long time horizon.

What artificial intelligence technologies are used in the work?

To build a language-independent semantic structure, technologies of syntactic and semantic analysis are used. Classification and machine learning technologies are used to resolve homonymy.

How much will your new technology cost?

We will try to make sure that everyone benefits from their use. So that translators can earn more by increasing productivity, so that the customer can translate more and pay less. Otherwise, you can’t do business in the modern world. You will be successful only if all your partners and clients are successful. Surely something will be available for free on the Internet.

Can (and will you?) Use Compreno technology to improve text recognition? I mean, will Compreno help make a decision in controversial cases with defects in a recognizable picture? Can Compreno work with an incomplete data set and help make assumptions about what should be in a poorly readable place?

Apparently some elements of Compreno technology will be used in the FineReader recognition system in the future.

How can your system help me, a simple builder?

Great question! You can take a photo of a bag of dry Chinese glue or a can of German paint on your mobile phone, click on the “Translation” button, the program will recognize and translate Chinese and German instructions into Russian, which will help you to properly spread glue or mix colors. You will do your job efficiently, and it will bring joy to those people who will live in the house you built. This is what we call "helping people better understand each other." We create artificial intelligence technologies that enhance the quality of life. In this sense, we are colleagues. You also create something that enhances the quality of life.

Continued next week.
Part 2. Grant
Part 3. Skolkovo, Linux and other issues