Guest | Wang Yan
Author | Yun Zhao
Column introduction: "T Frontline" is one of the in-depth interview columns specially opened by the 51CTO Content Center for technical figures. By inviting people in the technology industry Business leaders, senior architects, and senior technical experts provide in-depth interpretations and insights into current technology hot spots, technology practices, and technology trends to promote the dissemination and development of cutting-edge technologies.
Artificial intelligence is called the fourth industrial revolution and is used by society People from all walks of life have placed infinite expectations on it. It not only improves people's lives, but also improves the operational efficiency of the entire society. In recent years, large models and multi-modality have once again boosted the enthusiasm of related research.
In the spotlight, what everyone may see more is that the industry is pushing AI to the extreme, but often ignores its other side of "water conservancy for all things".
The development of technology is inseparable from actual scenarios. The many ups and downs of AI research craze are always accompanied by the problem of application landing. How to steadily push "high-end" technology to the ground? How do you view the deep integration of AI and industry segments? What do you think of the research craze on large models?
With these questions in mind, "T Frontline" interviewed a technology company that is deeply involved in the education scene: Zuoyebang. As a company committed to using technology to promote inclusive education, it uses cutting-edge technologies such as artificial intelligence and big data to provide a series of efficient services for hundreds of millions of C-side users and thousands of B-side customers. of learning and education solutions and products.
Especially in the field of education, Zuoyebang is a good representative of the exploration and implementation of AI technology, regardless of user size and technology selection. This also provides important reference for us to think about how to use artificial intelligence to promote industry changes. T Frontline is honored to invite Mr. Wang Yan, chief architect of Zuoyebang Intelligent Technology Laboratory, to provide us with his insights on these issues.
When it comes to homework help, what may be most impressive to you is the powerful question bank function. Zuoyebang is one of the first educational technology companies in the country to establish a question bank. So far, it has a question bank of 540 million. How was such a huge question bank constructed? According to Wang Yan, the success of question bank construction benefits from three conditions.
First of all, it stems from the innate advantages of Zuoyebang: As we all know, Zuoyebang was originally an internal incubation business of Baidu. It was initially positioned as a Q&A mutual aid community and later launched a search Q&A business. In order to optimize the search and answer results, Zuoyebang has established the largest online question bank production platform by forming a team of full-time teaching researchers and part-time teachers. This is also derived from a model that Baidu knows. In a community atmosphere that advocates sharing and communication, users are encouraged to solve problems with each other, and it is also very close to the actual Q&A and communication scenarios of netizens. Instead of what other companies did at that time: let part-time college students do the question bank. In this way, through in-depth analysis and mining of user-generated content, we gradually understand which problems users are most concerned about in the learning scenario, which problems are more difficult, and which problems most people will encounter. This is an important premise, which clarifies our construction direction.
Second, Zuoyebang attaches great importance to the value and construction of resources and pays enough attention to the question bank. The question bank not only plays a huge role in user communication, but is also a very important part in answering questions and teaching. Relying on the crowdsourcing system, the construction of the question bank is divided into independent process steps with less difficulty, making the question bank better and faster, and relatively comprehensive, which ensures the scale of the question bank construction.
Third, it is not enough to have one question after another. The questions also need to be related, such as: the knowledge points examined, the difficulty level, and other knowledge points they rely on are related to label terms. This involves the processing of tags and association with technical infrastructure such as knowledge graphs and knowledge trees. Only in this way can the question bank have the ability to be efficiently retrieved and filtered, so that the value of the question bank can be truly brought into play. Of course, many links in the construction process of the question bank itself were manually operated at the beginning. Later, AI technology was continuously introduced, such as taking photos of questions and most other electronic entry steps. AI automatically recognizes these images and turns them into computer-understandable formats. Data and language. Automated auxiliary processing is carried out through AI capabilities such as automatic labeling, formatting formulas, and AI error correction technology, which greatly improves the accuracy and greatly reduces labor costs. Thanks to the construction of the question bank and the continuous deepening and expansion of AI technology, Zuoyebang has implemented a series of AI acceleration technologies to optimize the response time of search and answer questions to 1 second, while the response time of early similar products was around 8 seconds.
In projects connecting public schools, the question bank plays a great role in assisting teaching scenarios. A highlight scenario is a high-quality homework system with the ability to predict questions accurately and personalized. The essence of the system is to conduct data analysis based on different student statuses, such as the time it takes to do the questions, and the mastery of different knowledge points, for personalized question recommendation. Because for students, questions that are too easy or too difficult will make the questions lose their value. The same question will have different values for different students. Therefore, it is necessary to have a full understanding of students, combined with the rich label dimensions of the question bank itself, and accurate matching. The question bank plays a great assistive role in the design of high-quality homework products, and is conducive to students consolidating the knowledge that really needs to be consolidated.
In terms of homework scenarios, in addition to the question bank, the more important thing is the automatic-assisted correction technology. Compared with objective questions, it is very difficult to correct subjective questions. Taking mathematics answer questions as an example, we use the OCR technology accumulated over many years to accurately identify the content of students' answers, and then use NLP technology to conduct structured analysis, such as logical analysis of the answer content, and then identify error points according to the answer specifications; In addition, the ability of knowledge graphs is also used to not only point out where students made mistakes, but also tell students why they made mistakes. It also uses the ability of user portraits and recommendation algorithms to generate student-specific learning reports to help students find weaknesses in the learning process. point. At the same time, relying on Zuoyebang's cloud-native and multi-cloud disaster recovery system provides high stability and reliability for this service system. Therefore, this ensures that even if many schools use it at the same time, there will be no downtime and stable use by users.
According to Wang Yan, due to years of accumulation of large-scale users of Zuoyebang, they regularly conduct performance evaluations of the operating system, and the evaluation results are also in an industry-leading position. Compared with similar products on the market, Zuoyebang currently supports more question types and has a higher accuracy rate.
1. Step-level automatic auxiliary correction
There is a closed loop in the learning of knowledge: the teacher imparts knowledge through teaching; Students test which knowledge points they have learned and which they do not know by doing exercises. The knowledge they do not know requires continuous learning and practice. In this closed teaching loop of "lecturing → doing questions → judging questions → speaking about questions", teachers repeatedly correcting a large amount of homework has become a major pain point. In the past, it was difficult for teachers to accurately tutor each student. Using AI to assist in correction can help teachers effectively reduce the burden of correction, significantly reduce unnecessary time and energy, and at the same time allow more students to effectively improve their grades.
Currently, the homework product system has a high usage rate, and teachers use it almost every day. Moreover, this system can also integrate the teacher's teaching experience and style and be customized according to the different needs of the teacher. Currently, subjective questions and applied questions can be graded based on step levels. The main direction for later improvements is to continue to reduce the proportion of teachers' manual work in correcting homework.
2. The answer to the graph question contains the universe
Compared with text questions, the correct questions can generally be identified and matched in the question bank through OCR, text search, etc. However, there is a special type of questions, which are graphics questions. For example, there are often questions like this on test papers: Find the shadow area of the following graphics. At this time, the features that need to be extracted are not just text, but also the features of the picture. Because only through text search, the search system behind the question bank can search for similar question stems, but the shape of each question in the results obtained is different. At this time we need to perform vectorized feature extraction on the image. The digital vector expression is aggregated with the characteristics of a large number of question banks to form the characteristics of "text images". Especially in elementary school questions, there are often many mixed images and text. It is not only necessary to understand the text of the question, but also to understand the structural relationship between the boxes, including the starting position of the features of the extracted lines in the "connection questions". trajectory to determine. The same goes for drawing questions.
3. Test paper restoration: black technology is often rooted in reality
In the context of auxiliary teaching, homework help It has also accumulated many technical patents in OCR, speech, image recognition, and homework correction. For example, we have disclosed a patent for artificial intelligence to efficiently correct distorted images, which was developed in a very demanding scenario. As we all know, "redoing wrong questions" is a very important part of the teaching scene. Parents and teachers need to restore the test paper to the unanswered state. However, after taking pictures of the test paper, the handwriting often appears uneven and the questions on the test paper are distorted. Therefore, technology is needed to solve the problem of typesetting correction.
We use deep neural networks to identify human handwriting and distinguish it from test paper fonts. Combined with image enhancement technology, we can restore test papers very effectively. Currently, this The technology has been launched in the Zuoyebang App and has been applied to printer products, which can restore paper test papers to their original state. It only takes a few seconds from taking a photo to restoring it. In the previous practice, students usually needed to copy the questions manually and then do it again. This can be said to be a powerful "black technology". Of course, this technology is not only used to restore test papers, but can also be used to correct and beautify the photographed homework before submitting the homework in online classes to restore it to a better layout state. On the one hand, it is conducive to preservation, and on the other hand, it can also improve Content recognition accuracy.
4. Knowledge graph: a gathering place for expert knowledge
The construction of knowledge graph is inseparable from human experience The same is true for the system and the educational scene. Our knowledge graph capabilities are more accumulated in course scenarios. A large number of teaching and research teachers summarize the context, dependencies and learning paths of knowledge points during the teaching and research process. These relationships and paths can connect scattered knowledge points into a network, thus forming the prototype of a knowledge graph. Teaching and research teachers provide a wealth of expert experience and knowledge point systems. In this process, the R&D department uses a series of automated AI machine learning capabilities for large-scale implementation. After having the map, we can make the next step of personalized homework design, such as recommending questions that are equivalent to ability or even challenging to learn deeper knowledge points. At present, knowledge graphs are used in a wide range of application scenarios in homework help: teaching scenarios, homework correction, personalized learning, homework diagnosis, including the correlation of relevant questions in the question bank just introduced, which essentially gives the questions a more precise dimension for retrieval and recommend.
In the past teaching scenarios, on the one hand, paper books, teachers’ blackboard writing, PPT, etc. were not used To digitize, on the other hand, students’ response content, including whether the answers are correct or not, homework and test scores, etc. also need to be digitized. Why digitize? Because if the content of the natural physical world is not transformed into data information that computers can understand, our advanced technology research in the computer field will not only be unable to be implemented, but even technologies that only improve efficiency, such as retrieval and recommendation, will be useless. Therefore, both speech and images are important media and carriers for conveying teaching ideas and knowledge, and these require in-depth digitization.
In recent years, with the continuous advancement of education informatization, most classrooms have been equipped with large digital screens. The teaching courseware used in daily classes has been digitized, and what we are doing now is to promote homework The scene is digitized. However, it is worth noting that the ability to use AI at this stage must respect the original habits of teachers and students and should not be changed easily. For example, in the original teaching model, everyone is accustomed to paper test papers. If you cancel the paper test papers and move them all online, there will be a serious problem of "acclimatization". Although online answers are required to be digitized, this changes habits. Once you change your habits, it becomes difficult to use them on a large scale.
Based on this, out of respect for teachers’ real habits of marking papers and students’ answering, Zuoyebang has innovated its business ideas: it has introduced the function of “leaving traces of original papers” in the homework system. Therefore, in Wang Yan's view, what we need more is to innovate in thinking, to lower the threshold for using technology, and to digitize without changing habits.
When you zoom in from the workplace scenario to the education scenario, you will find many new needs in new scenarios. For example, in the sports scene: During class, the physical education teacher pays great attention to the exercise intensity that each student can bear, such as heart rate monitoring. When a student's heart rate is too high while exercising, the student should be reminded to stop and rest. Another example is "skipping rope counting". We don't use a counter, but it would be more convenient to let the camera automatically identify and count. In addition, the capture of body movements is also a practical technology to help students check whether their movements are standard and standardized. These AI can assist in correction.
1. How to find implementation opportunities for AI
Zuoyebang is a technology-driven company. Development teams often ask questions like: What other technologies could be used? Is there any good technology that can make it possible to meet the needs that were not met in the past, and to complete the things that were previously unachievable? Based on this, Wang Yan summarized the logic behind how to find implementation opportunities for AI: We should know what technologies we have mastered and what resources we have, and then consider how to apply appropriate technologies to specific scenarios. Based on existing technical resources, perform scene matching. The next step is to think and weigh what the technology can achieve, and then pilot and optimize it.
2. B-side accuracy requirements are more stringent
In the operation scenario, compared with the C-side scenario, B-end customers have special needs and have customized requirements. For example, schools will have higher requirements in terms of accuracy, and in the correction process, mistakes must not be made. While C-end products focus more on the richness of functions and user experience, the expectations for accuracy are not that extreme.
1. Basic research is the foundation, and cutting-edge technology is more fragrant
Basic research is our technical base. These technologies already have a wide range of application scenarios. Optimization of basic technologies will bring about considerable improvements in application performance, so investment in basic research is essential. of. The research on cutting-edge technology may bring about changes in gameplay. As technology continues to develop and innovate, what could not be done before may suddenly be possible one day. Students in the laboratory are encouraged to allocate 20% to 30% of their energy to Pay attention and follow up. In terms of candidate abilities, we hope that while having certain academic research abilities, we will also value engineering abilities. Stronger engineering capabilities mean stronger implementation capabilities, and to truly promote the implementation of AI technology, we need to continue to be solid at the application level. Of course, ideally, we hope that talents have full-stack capabilities and can independently complete experimental design and application implementation to quickly verify the improvement effects of certain innovations in practical applications.
2. The model cannot simply pursue bigness, but the benefits must also be broad
The practice and development experience in the education field and general industry are still somewhat different from general technology. AI has now been applied to all walks of life, but when it comes to education scenarios, most models based on general scenarios cannot be "generally beneficial" to specific scenarios. Wang Yan gave us an vivid example, such as handwriting recognition in handwriting input method. The algorithm model has an assumption: handwriting written by adults. However, in homework scenarios, students of different ages write differently, and there are not so many requirements for neatness and neatness. Therefore, for the education field, AI needs to be refined based on specific scenarios. It needs to sink into specific scenarios to settle down, solve the problems that are not good in general fields, explore and discover new business needs, and solve practical problems. in the process to promote the development of related technologies.
What is widely used must be affordable to the public. "Large models have achieved recognized performance improvements, but they are still far from being widely used by users." In Wang Yan's view, large model and multi-modal research can indeed bring about improvements in accuracy. But often when the accuracy of a task is improved, such as from 95 to 96, this point of improvement comes at the cost of a huge sacrifice of computing power. Nowadays, large models with hundreds of billions and trillions of parameters require very large clusters to support and run them. For actual scenarios, if there is no huge computing power cluster to use, the results of large models can be generated within one second. , which may require quite a long running time. Although the performance of cluster hardware is constantly improving and the corresponding unit computing power cost is constantly decreasing, technology that can be widely used must be low-cost and affordable. To a certain extent, focusing all your energy on the pursuit of computing power is a bit of a sacrifice. The widespread implementation of AI lies more in the innovation of ideas and the pursuit of cost-effectiveness of the technology itself.
How to benefit technology to thousands of users and how to widely use gimmick-like functions in real life is a key issue. We currently have a very large number of users, and there are a lot of users using it every second, so if we do it by "piling up computing power", the cost will be unimaginable. At present, under the circumstances of affordability, what we have to do is to provide users with as rich functions and services as possible. On the one hand, we think about how to improve the utilization of computing power so that the equipment is not idle. On the other hand, we explore how to improve and optimize the model and engineering architecture to provide tens of thousands of operations per second at the most reasonable cost. level of large-scale AI services. More importantly, how to innovate ideas. Only by innovating from the perspective of solving problems and allowing more people and users to actually touch and feel the convenience brought by technology can greater value be released.
Wang Yan, chief architect of Zuoyebang and head of Zuoyebang Intelligent Technology Laboratory. He once served as the technical director of Baidu Zhizhi and Baidu Encyclopedia. He currently serves as the head of the Intelligent Technology Laboratory of Zuoyebang. He focuses on related research and implementation in the fields of artificial intelligence, image technology, large-scale high-concurrency online architecture and other technical fields. He is mainly responsible for Zuoyebang. Search Q&A, AI correction, question bank and other related services.
The above is the detailed content of T Frontline | Exclusive interview with Wang Yan, chief AI architect of Zuoyebang: The inclusiveness of AI lies in innovative ideas. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
