Research on Content Understanding Algorithm

After several years of expansion, the popularity of algorithms has declined rapidly, whether it is the difficult listing of AI Four Little Dragons, the return of vice presidents of major Internet companies to academia, or the freezing of recruitment of algorithmic personnel. There is also the impact of the deterioration of the overall economic situation, which is also related to the upper limit of the algorithm's own ability. In all kinds of learning tasks, the performance of the algorithm gradually enters the bottleneck, and the gradient of improving the effect of general tasks gradually decreases. Effective progress depends on very large-scale data and model parameters. Taking the general semantic representation task as an example, the cost of completing a super-large-scale pre-training model reaches several million yuan, which greatly limits the opportunities for SMEs to participate.

In terms of business application, after several years of continuous construction and output of various functions, the content understanding algorithms corresponding to different business dependence directions have been relatively mature in application and effect, and there are few opportunities for surprises. In this case, as a part of background support, how to identify and deepen the role of content understanding algorithm becomes very important. This paper attempts to analyze the living conditions of content understanding from the perspective of value, explore the possibility of future development, and make some less rigorous explanations for employees' coping styles.

We have always defined the content understanding algorithm as a panacea for business, which can be plugged in anytime and anywhere. Cooperate with content producers to give creative tips, help operators do quality analysis, copyright protection and similar search, help search algorithms improve the efficiency of long-tail query, help recommendation algorithms provide fine-grained semantic features such as labels, and block the same type of content (such as soft porn, disgusting and disliked stars) according to negative feedback from consumers. Therefore, we naturally define the mission of the content understanding algorithm as "improving the quality and efficiency of the whole link of the content stream", in which the quality includes the removal of deterministic inferior quality and the trial-free or high-exposure recommendation of high-quality content. Efficiency refers to optimizing the time from production to consumption of content to the fastest, including rapid content screening and accurate matching of people and content with auxiliary distribution algorithms.

What needs to be answered here is, on the basis of the above relatively complete capabilities, what is the core value that the content understanding algorithm can provide?

The first is the definition of customer. The customers of content understanding algorithms are not operators, distribution algorithms, producers or consumers. It is to return to the original "content" and maximize the added value of the content.

Secondly, the roles involved in each link of content circulation bear the value of content understanding algorithm, whether it is the realization of platform will led by operation, the efficient matching of distribution algorithm to content and consumers, or the demands of producers and consumers for content understanding algorithm from content supply and consumption respectively.

Finally, the content itself is a carrier, behind which people describe the real world. With the logic of accessing content on demand at the platform level, consumers also have the right to vote with their feet. The content understanding algorithm here should not be regarded as any self-inclination, and the core value of content understanding is the ability to output according to the diversified business demands. Otherwise, the open source model simply moved to academia can pile up superficial business capabilities, which obviously cannot meet the demands of business growth.

Therefore, we can form a value definition: "The core value of content understanding algorithm is to provide intelligent and structured understanding ability according to the diversified demands of service business in the whole life cycle of content, and its measurement standard is the additional efficiency improvement and cost reduction brought by the above ability."

From this definition, it seems that the content understanding algorithm does not stand on the main battle line, and its value is implicitly calculated. Actually, it is not. Just like a war, there are only a few troops in the front, and the arms that undertake defense and auxiliary tasks are actually indispensable, which often determines the direction of the war. There are many examples in history that the direction of war has changed because of logistics support problems. Just like the efficiency improvement in the definition, it is not a one-vote method. Because the improvement of efficiency will increase the scale of suppliers and consumers, it will also generate more demand for the efficiency improvement of content understanding algorithms. This positive feedback link is also an important way for content services to quickly achieve the set goals.

First, the valuation game.

Judging from the valuation of Xiaohongshu 202 1, 1 1, the content community with 60 million DAU and 47 minutes per capita was recognized by the valuation of 20 billion dollars, which was the height that an early content community could reach after a long period of growth. Considering a relatively innovative content business, it is actually very difficult to reach 50 million DAU in two to three years, with a per capita average of 5 minutes. According to the logic of benchmarking little red book, the upper limit of valuation is $3 billion. Assuming that the contribution rate of content understanding algorithm to business is converted at 3%, the estimated value of content understanding is $90 million. According to the simple marketing rate of 10 times, the content understands that the annual revenue is 9 million US dollars (according to the valuation,

$9 million is an embarrassing figure because the cost of content understanding algorithm is relatively high. From a big perspective, the content understanding algorithm expenditure is divided into three parts. The first part is the algorithmic staff, which is calculated by the support team of 15 people (it seems a bit much, but it is not enough to support the expected growth in the future, and 15 people is actually not enough). According to a person's one-year payment cost, it is 65438+. The second part is resource consumption. According to the scale of millions of content/day, the cost of various resources (including machines, storage, auxiliary software, etc.). ) about $5 million/year; The third part is support engineers, product managers, outsourcing labeling support, etc. This part is about 654.38+500,000 USD/year. It can be seen that at this stage of business development, the content understanding algorithm is not enough.

According to the above caliber, the core methods to improve the value of content understanding include three directions. First of all, the increase of enterprise value requires the steady growth of DAU and duration. Second, the business contribution rate of content understanding algorithms has increased. The requirements for content understanding algorithms here are relatively high, not only from the quality of content, the assistance of producers, the efficiency of distributing traffic, the overall ecology of business, but also the commercialization of business. The third is to reduce costs. The feasibility of this road is weak. On the contrary, with the growth of business, the consumption of costs will further increase. What we can do is to control the growth rate of cost below the growth rate of business.

It is not optimistic to measure the value of content understanding algorithm in the early stage of business development according to business contribution.

Second, value remodeling.

The former perspective is the value measurement under closed-loop business, which is open to a larger perspective. The reason why the content understanding algorithm constitutes a relatively independent functional unit is that the capabilities it provides are relatively universal. For example, the tag recognition algorithm can be used not only for small red books, but also for content services such as Tik Tok and A Auto Fast.

Therefore, in the process of supporting specific services, it is another door of content understanding algorithm to precipitate the value of a general algorithm to output other similar services. Another problem faced here is that if it is a head business, it must require the content understanding algorithm to be tailored for it, and the price that small and medium-sized businesses are willing to pay for the content understanding algorithm is limited. The core solution to this problem is to provide the most universal ability when customizing the head business ability, and realize the stacking of orders of magnitude by forming value output for a large number of small and medium-sized businesses.

In addition, in order to distinguish mature business from innovative business, for mature business, the slight improvement of content understanding algorithm may be very obvious. Taking JD.COM platform as an example, if the content understanding algorithm improves the business transaction by 0. 1% through image search or paragraph recognition, it will also increase the value by hundreds of millions of yuan every year. For innovative business, the content understanding algorithm should go deep into the business, provide hard-core ability for the business from the whole life cycle of the content, help the business realize the obvious improvement of the physical sense of producers and consumers, and finally bring positive feedback growth to the business. In the early days, Tik Tok relied on the cool AI special effects system to realize the rapid growth of user scale.

"At present, the effective way to amplify the value of content understanding algorithm is to precipitate the generalization ability and output as many services of the same type as possible in the process of personally serving the head business. Find some growth points for business growth for mature businesses, and innovative businesses find hard-core capabilities suitable for rapid business growth. "

As a veteran of content understanding algorithm for six years, I am cautiously optimistic about the future of content understanding algorithm. There are three reasons. First, the space for algorithm improvement brought by this round of deep learning becomes limited; Second, after Internet users enter the stock age, the content of the head will be more intensive, from the pursuit of efficiency to the refined operation and sticky maintenance; The third is to look forward to the next generation of content consumption patterns that may appear in the future.

First, the space for algorithm improvement is relatively limited.

In the past few years, the evolution of content understanding algorithms can be divided into three directions. One is the upgrade from traditional manual features to neural network features. Through big data and computing power, the effect is obviously improved, and the entry threshold for algorithmic personnel is greatly reduced. Second, the understanding of content has been upgraded from a single mode to a multi-mode &; Cross-modal reasoning ability based on graphic neural network: Thirdly, the model learning of extremely large-scale data, that is, the unified content representation method based on large-scale pre-training model, has given birth to the growing transformers family.

However, the performance of the algorithm is gradually approaching the bottleneck, and there is still a certain distance from human beings in the algorithm tasks such as image reading, sentiment analysis and tag recognition, and there seems to be no clear breakthrough opportunity in this distance in the short term. On the contrary, the industry began to move from supervised learning to unsupervised learning, trying to use massive data to learn the paradigm behind it, which is essentially deviating from the ability to catch up with human beings.

Take the transformer as an example. The training of tens of billions of data consumes hundreds of GPUs at a time, and it takes several weeks of training time to improve the effect obviously, which does not include the frustrating time cost of fine-tuning the network. In addition, if the downstream tasks want to get the expected results, they need further transfer learning. From the appearance, it only provides a better starting point for algorithm learning.

We have experienced a process of cooperation, in which the business is waiting for us to become more rational. AI algorithm has never been a savior, but a more productive tool. Of course, we should not be too pessimistic. At least in the past few years, the booming algorithm system has brought about a great decline in the entry threshold for employees. The public's extensive understanding of AI algorithm has also contributed to the relatively long-term vitality and growth of content understanding algorithm.

Second, the operation mode of content community in the era of stock users

The use of Internet in China means that major content providers must enter the stage of stock users. The dilemma faced by stock users is that extensive growth no longer appears, user groups begin to be subdivided, user stickiness becomes more difficult, and content communities need to be refined. The performance behind refined operation is that the demand for efficiency drops, and it is replaced by patience with users' minds and long-term tactics. In this case, the content understanding algorithm will become a functional support point scattered in many business requirements lists, and the chance of independence is decreasing.

"From the perspective of algorithm learning, people's creativity, gameplay design and interactive attributes are the real ceilings on the ground, so it is relatively reasonable to keep the tool attributes at this moment."

Third, the next generation of content consumption patterns.

Content consumption in the Internet era has experienced an upgrade from words to images to videos. The output behind every content consumption upgrade has exploded into a content understanding algorithm. So what is the content consumption pattern of the next generation?

The industry is currently betting on the Metauniverse, and facebook even changed its name to Meta. There have been several waves of VR/AR craze in the past. It seems that apart from some online adult websites and offline game devices, there is not enough output to subvert our daily life style.

Obviously, human beings need to perceive the external environment at a higher level and interact with others without time and space differences, but whether it is carried by the "meta-universe" is unknown. If the metauniverse is used as the carrier, emotion recognition, tactile sensation generation, natural interaction, ecological health management and load reduction under the super-large-scale content consumption in the virtual world will be a brand-new zone that the content understanding algorithm can try to conquer and deepen, and will also assume a more core role.

"The consumption pattern of the next generation of content understanding has the opportunity to become the next main battlefield of content understanding, but the current situation is not clear, so we need to keep patient thinking and wait and see".

Fourth, other possibilities.

Aside from the head comprehensive AI manufacturers such as Baidu, Tencent, Alibaba, Huawei and other enterprises as the first pole of content understanding diversity demand output, there are also content understanding algorithms as the second pole of platform capability output. The famous AI Little Dragons (Shang Tang, Defiance, Yi Tu, Congyun) and industrial AI capability output that deeply integrates various fields of people's livelihood.

Medical AI has solved the problems that medical resources are insufficient to meet the demand for medical treatment, and manual consultation takes a long time. A typical case is coronavirus pneumonia -Moonshot crowdsourcing protocol, which involves more than 500 international scientists and aims to accelerate the research and development of coronavirus pneumonia-19 antiviral drugs.

Education AI solves the unfair distribution caused by the lack of high-quality educational resources and asymmetric information between teachers and students. Although the country is pushing the new policy of "double reduction in education", education, as a basic individual right, should be better satisfied. Well-known enterprises include squirrel AI and ape tutoring.

Manufacturing AI has solved the problems such as the increase of equipment, quantity and function, the difficulty of dispatching and distribution, and the personalized demand of demand side. By using AI, automation, IOT, edge computing, cloud, 5G and other means, we can make full use of the massive value data in the production workshop, free people from simple and repetitive labor, and engage in higher-level tasks, helping to increase production while reducing the defective rate. Some well-known enterprises are innovative and innovative industrial artificial intelligence listed in Hong Kong.

In addition, there are various companies that have been working hard in industries such as smart driving, smart cities, and chip AI. They are giving full play to the capabilities of big data and AI algorithms and bringing endless innovation capabilities to major industries.

Back to the existing living environment of content understanding algorithm, there are still some potentials to be tapped. Before the next generation of content consumption comes, we can do better, form a benign linkage with the upstream and downstream, and show a better style on the current stage.

I. Products

Content understanding algorithm Whether this product is just needed is a bit controversial. Some people say that the output speed of the algorithm is slow, and it will be a waste to let expensive product roles participate in the construction itself. Personally, I think the product role corresponding to the content understanding algorithm must be possessed, because behind the huge business system, if there is no top-down design and construction of the content understanding algorithm system for business needs, it is very easy to get used to business empowerment.

The core problem that products need to consider is how to measure long-term and short-term investment. Algorithm is a delicate work, and the expectation of the result is uncertain. Therefore, it is necessary to manage business expectations and interact with business requirements in a timely manner. In order to ensure the final use effect of the algorithm in business, we can make quick trial and error by simplifying the version of the semi-finished algorithm or the product scheme in the early stage, which is helpful for business decision-making and wins space for the long-term iteration of the algorithm. In addition, designing an effective sample data return mechanism for the long-term iteration of the algorithm, outputting as many trial and error methods as possible to the business through configuration, and monitoring the effect of the business in real time are all work that the product needs to think about.

Second, the operation

Operation should be the most frequent participant in the content understanding algorithm, and the evaluation criteria and business adaptation of the content understanding algorithm need to be constructed and monitored by operation. Content understanding algorithm is an intelligent assistant for the operation of content supply ecology and consumption ecology. It provides and operates various analysis and use methods from the perspective of content structured tags, such as content review, content circle selection, content crowd fixed investment and so on.

Processing operations put forward high requirements for content understanding algorithms, and how to quickly measure the rationality and feasibility of the requirements is very critical. Sometimes the content understanding algorithm is over-invested, which leads to poor online effect and affects business development. Sometimes the lack of confidence in the implementation effect of the algorithm, or the lack of leverage in production and use, leads to the rejection of demand, which leads to the loss of valuable trial and error opportunities for the business. Therefore, the content understanding algorithm should grasp the links of content operation well, define the full link algorithm capability together with the operation, and promote the reasonable and orderly development of algorithm requirements from the perspective of application.

Third, producers.

Producers are very important to the platform, and it is difficult for a clever woman to cook without rice. No matter how powerful the operation and distribution algorithm is, without high-quality content production sources, the business cannot continue to grow. Under normal circumstances, 2,000 high-quality producers plus tens of thousands of ordinary producers can support tens of millions of DAU businesses. How to serve these producers well is very important for the platform.

At present, the interaction between content understanding algorithms and producers mainly includes several aspects. One is the intelligent recommendation of the producer's content elements in the process of content production, such as topic, title, soundtrack recommendation and so on. The second is to improve the content display effect, such as filters, stickers, beauty, image quality enhancement and so on. The third is to guide and control producers from the quality level, including the reasons why the content released to producers from the commercial point of view is not adopted by the platform, the consumption list of high-fever content, and the copyright protection of content.

From the producer's point of view, it is the fundamental pursuit to get as much traffic or commercial value as possible from the platform, so there are often constant exploration of platform rules to gain benefits, such as releasing a lot of marginal balls or alarmist content. The content understanding algorithm needs to help the platform maintain a healthy ecology and effective traffic distribution, and give producers as much guidance as possible when the content supply scale is increasing. This killing relationship also brings a lot of challenges and living space to the content understanding algorithm.

Fourthly, allocation algorithm and consumers.

The core logic of putting distribution algorithms and consumers together is that content understanding algorithms need to deal with consumers through distribution algorithms in most cases. From the consumer's point of view, highly active users represent the mainstream mind, and how to serve this group well is related to the life and death of the business. Low-middle active users are the increment of the platform, and it is a key task to continuously strengthen the platform stickiness of these users (some users will flee here, and this part of sacrifice is acceptable in order to maintain the platform's mind). The distribution algorithm undertakes the mission of recommending massive content according to users' long-term and short-term interests after load reduction. The distribution algorithm needs to adhere to the will of the platform and be used for content traffic distribution to influence consumers' body feeling and mind, bringing endless vitality to the platform.

In the early editing-oriented content distribution model, consumers were the objects of education, and they could see very little fresh content in one day, which led to the limited browsing depth and duration of consumers. In the personalized recommendation mode, users' interest is greatly enlarged, and consumers will feel a strong sense of immersive consumption due to the rapid push of related content and fresh content. However, the diversity of content, the continuous maintenance of consumer perception and the expansion of interest have become very important, which puts high demands on the accuracy of the distribution algorithm. Providing the fine-grained recognition ability of distribution algorithms is an opportunity for content understanding algorithms to display their talents. Does the content have a good distribution potential to increase distribution traffic? What audience does the content suit? What are the core interests behind users' disorderly browsing? Soft porn/how to accurately identify the content that some people don't like (snake and insect pets) for distribution supervision and other issues are difficult for distribution algorithms to touch. These propositions are important aspects that content understanding algorithms can deeply study and affect content distribution and consumption.

Except for certain scenes (such as interactive play, personalized cover map, etc.). ), the content understanding algorithm should abide by its participation breadth in the content life cycle. When it comes to the distribution and consumption of content, the content understanding algorithm should define itself as an indispensable auxiliary tool of the distribution algorithm, rather than trying to replace it. From the perspective of content understanding algorithm, the distribution algorithm can be approximately equal to the consumer. Taking restaurant operation as an example, the delivery algorithm is chef, which provides personalized catering services according to consumers' tastes, ingredients and recipes. Content understanding algorithm can control the quality of ingredients, develop new recipes and provide semi-finished dishes when necessary. Interaction with consumers is handled by the distribution algorithm. After all, there is specialization in the industry, and the content understanding algorithm can make a vertical depth in the deep understanding of content and consumer insight, providing more possibilities, including ecology, diversity, content preservation and so on.

The ideal state of content understanding algorithm and distribution algorithm is positive sum game, and zero sum game is meaningless to both parties. Therefore, the additional requirement for the content understanding algorithm here is to establish a relatively objective evaluation system in the context of content consumption, accelerate the online process through standardized evaluation of the algorithm, and provide more guns and ammunition for the distribution algorithm through continuous rapid trial and error.

Verb (abbreviation for verb) Project &; data analysis

A hero has three gangs, and a group of friends stand behind the content understanding algorithm. The large-scale engineering infrastructure of algorithm production and the data analysis ability of algorithm insight can help the content understanding algorithm to develop better. In today's explosive growth of content, an efficient algorithm engineering system is very critical, even one of the most important means to open the gap between different companies. There is an obvious example. In all kinds of algorithm competitions held by the industry, as long as large Internet companies participate, they will basically dominate the list, which is behind the strong first-Mover advantage of model training efficiency. There are very few universities with 100 GPU concurrent training ability. In addition, taking the general vector retrieval function as an example, a large number of engineering optimization methods are needed to stably run tens of billions of vector indexing capabilities with limited computing power and memory consumption, and this function is very important for the efficient use of the algorithm.

Data analysis has many applications for content understanding algorithms. According to the statistical behavior of consumption, the user portrait oriented to content interest is constructed, which provides the wind direction and trend direction of content consumption, the reasonable subordinate relationship of layered coupling content tags, the effectiveness analysis of the algorithm before it goes online, the continuous monitoring and abnormal alarm after the algorithm goes online.

What the content understanding algorithm needs to do is to design a complete architecture for the business field. From the perspective of algorithm efficiency, it includes the linkage project of algorithm service efficiency and algorithm insight perspective. Data analysis provides powerful productivity, and builds enough technical threshold through scale and system thickness.

202 1 This year is different for China Internet and even China society. With the global economic downturn and the construction of artificial barriers between countries, the country has just experienced the stagnation of domestic population growth, the strong control of the Internet platform by the state, and the peak of Internet users' penetration.

The AI algorithm system behind the content understanding algorithm has also suffered some twists and turns, but overall, the AI algorithm system and industrialization are still developing. In the basic theory R&D system, the number of published papers, conferences, competitions and participants has increased. The cold weather in these two years is mainly affected by the market environment, and the proportion of AI in total investment is still rising. Domestically, the technological blockade of the global economy has further strengthened China's determination and confidence in independent innovation. China's "14th Five-Year Plan" has clearly put forward the industrial development plan of big data, artificial intelligence and VR/AR, and the AI industry still has strong potential worth tapping.

As a business-dependent content understanding algorithm, it needs to have clear self-awareness and positioning. What are our core values? How to effectively define and quantify? As one of the many roles that serve the business, how to interact with other roles well? The answers behind the above questions represent the core role of content understanding algorithms. At present, there is a speculative wind in the field of algorithms, and we do whatever is popular, such as claiming unsupervised learning without data labeling, claiming that it can effectively learn from a large number of labeled samples, and claiming that a single algorithm model can be multi-modal pre-training learning all over the world. There is no problem if we study the basic theoretical system and algorithm learning paradigm from the perspective of abstract simplification of problems, but there is a problem if business students also talk about such concepts. It is arrogant to customers to talk about technology to create new business without the core requirements of business scenarios, which is a typical mechanism.

As a content understanding algorithm for deeply converged services, we should explore the core technology of business empowerment based on the feasibility of business scenarios and algorithms. Even if it takes a long time to polish the algorithm technology that can produce positive commercial value, we should dare to invest in the construction, constantly think about more possibilities of business in the process of algorithm research and development, and gradually turn the commercial uncertainty into the relative uncertainty of technology. For algorithms that cannot contribute to business for a long time, we must resolutely give up in-depth research. Of course, as a technical follow-up, it is no problem.

Judging from the current situation, the development of content understanding algorithms has indeed encountered some difficulties, but we can be cautiously optimistic about the future and look forward to the arrival of the next generation of content consumption patterns. At the same time, it is necessary to refine the business scenario as much as possible to output and strengthen the ability, amplify the commercial value of existing stocks, and make technical reserves for possible moments in the future through the continuous construction of the algorithm itself.