Ant Financial: How artificial intelligence is "calculating carefully" behind the Double Eleven

Double Eleven, Carnival. Overnight, Taobao and Tmall users spent hundreds of billions, with a peak number of online users of tens of millions per second and a peak transaction volume of tens of thousands per second. Every purchase by a user affects every nerve of the merchant from stocking, dispatching, warehousing, logistics to after-sales service.

Artificial intelligence has become the "invisible hand" behind the extremely high load requests. In terms of customer service alone, on November 11, 2015, Taobao + Tmall answered more than 5 million user questions through self-service, increasing Ant Financial's customer service efficiency by 20 times.

This year, we invited Li Yong, the head of Ant Financial's intelligent service platform, to explain to us:

• What challenges does Ant Financial face during Double Eleven?

• Where will artificial intelligence play a role in the Double Eleven?

• How to combine the research and development results of the artificial intelligence department with Alibaba products to serve a super-large user base

• Features and difficulties of Ant Financial’s existing major intelligent services in practice

Introduction of Li Yong:

He is currently the head of Ant Financial’s intelligent service platform.

He joined Ant Financial in 2015 and is mainly responsible for the planning and implementation of intelligent services, intelligent operations, and operation-related technical products.

He used to work as a senior R&D manager at ask.com, and has been engaged in the R&D management of search-related technologies for a long time. He has rich experience in R&D management of products and technologies related to data, search, personalization, and traffic.

The following is the sharing content of this open class, compiled by Leifeng.com (public account: Leifeng.com).

1. Overview of Ant Financial’s Artificial Intelligence Department

Briefly introduce Ant Financial's artificial intelligence department and its role

The leader of Ant's Artificial Intelligence Department is Qi Yuan, who holds a Ph.D. from the MIT Media Lab and is a tenured professor in the Department of Computer Science at Purdue University. He has served as the executive editor of the authoritative machine learning journal Journal of Machine Learning Research and the field chair of ICML, the world's top machine learning conference. He has won the Microsoft Newton Research Award and the NSF Career Award.

Since artificial intelligence is based on multiple disciplines, most of the talents in this team come from large companies and universities. Many of them have returned from Silicon Valley and have many years of R&D experience in the field of artificial intelligence. They are a very professional and elite team.

(We’re sorry, but we can’t reveal the size and structure of this team yet).

This team provides artificial intelligence support for the products and platforms of the company's various business groups, supporting various businesses such as risk control, credit decision-making, insurance pricing, service recommendations, customer service, etc.

What are the points of integration between artificial intelligence technology and Ant Financial’s front-end, middle-end and back-end?

The most mature and typical R&D is intelligent customer service. After it was launched, it saved the company 100 million yuan after deducting the cost. This mainly refers to the labor cost. Just a few hours ago, the self-service rate of our intelligent customer service broke the record again, reaching 99%. I will talk about this in detail when I talk about intelligent customer service.

Let's take the example of return shipping insurance. This is the most familiar insurance. Some people may ask, if the buyer returns the goods frequently and at will after buying this insurance, won't the insurance company lose a lot of money? But in fact, the price of this insurance is different for each person, which is what we often say.

Behind this is the fact that Alipay's artificial intelligence comprehensively calculates the return probability of each order, taking into account factors such as the consumer's gender, shopping habits, shopping categories, selected stores, etc. For example, female shoppers are more likely to return goods, and the return probability of clothing categories is higher than that of beverage categories. Artificial intelligence learns a large number of models and a large number of rules. Insurance companies refer to the judgment of artificial intelligence and customize the freight insurance coverage for each person.

Let's take another example. After the launch of the "Airline Ticket Refund Insurance" jointly developed by Ant Financial and insurance companies, the claim ratio was as high as 190%, and the insurance companies were facing huge pressure of losses. However, by introducing machine learning technology and optimizing big data modeling, the claim ratio was effectively reduced, and the loss was successfully turned into profit, meeting the underwriting requirements of the insurance companies.

Among the key application directions of artificial intelligence in finance, we have already achieved a relatively high level of integration in customer service. What is our current research/engineering focus?

In the field of financial life, we have built an artificial intelligence platform and applied artificial intelligence to various service scenarios, such as risk control, credit decision-making, insurance pricing, service recommendations, customer service, etc. By using the power of artificial intelligence, we have greatly expanded our service scope, reduced risks, improved user experience, and cut costs.

When we talk about inclusive finance, a big challenge is how to use limited human and financial resources to provide the widest and most personalized financial services to the largest number of users. Ant Financial's vision is to provide inclusive financial life services to more than 2 billion users around the world in the next ten years. Only by leveraging the intelligence of machines can we solve the inclusive problem at this scale.

When we talk about machine intelligence, there are one or two very basic tests to determine whether a machine is truly intelligent, such as whether it can pass the Turing test or the Feigenbaum test. We hope to create machine experts in the financial life field in the future. When you talk to a machine, you don't feel that it is a machine, but rather that it is as professional and caring as a financial service consultant at a bank.

What challenges does Ant Financial have to face during Double Eleven?

Double Eleven has been going on for eight years and has now become a normal event. The payment guarantee capability that everyone is most concerned about is no longer a big problem.

During this year's Double Eleven, Alipay's core transactions, payments, memberships, accounts and other core data chains all ran on Oceanbase, a database independently developed by Ant. Ant's independently developed elastic architecture can utilize cloud computing resources in many cities across the country and easily hold this year's payment peak of 120,000 transactions per second. This new record is 1.4 times last year's peak.

In order to accommodate global payments and address time differences, the group promotion has two parts. Purchases by buyers mainly from China and Southeast Asia on Tmall and Taobao will start at 0:00 on the 11th Beijing time; purchases by buyers mainly from other countries and regions on Aliexpress will start at 0:00 on the 11th US West Coast time.

Although overseas banks and payment channels are widely distributed, they have experienced few major promotions and have limited ability to withstand sudden peaks during major promotions. To this end, this year, international payments have optimized the system and established a strong flood storage and buffer queue capability in the Alipay system, combined with user experience, to minimize peaks and fill valleys as much as possible to protect international card issuers and channels. Before the major promotion, an online real capacity survey was also conducted on key channels and card issuers.

This year on Double Eleven, there were less than 30 students in the Ant Platform Data Technology Group who were all in on Double Eleven technical support. You should know that our technical team has more than 3,000 people.

Since 2010, we have set ourselves the goal of reducing the annual investment costs specifically for Double Eleven security by about 30% to 50% compared to the previous year. By 2019, we will no longer spend a penny on technical support for Double Eleven.

In terms of payment security, how great has the pressure been in this area in previous years?

Three or four years ago, every time there was a big promotion on Double 11, we would mobilize 70-80% of the company's technical staff to all-in on technical support. For example, on the day of the big promotion in 2010, hundreds of technical staff (now the number has risen to more than 3,000) all sat in front of their computers, staring at the computers and checking the system. If there was a problem, they would immediately respond to it. If it didn't work, they would restart the machine, and if the capacity was insufficient, they would add a machine. We jokingly called this "human cloud computing."

Near the early morning, everyone thought that they had finally made it through, but we were hit by another blow. The peak suddenly appeared in the last hour, and there was a capacity problem with the data. At that time, there was still half an hour before November 12. What should we do? Fortunately, when designing the accounting architecture, we left ourselves a downgrade capability. We stopped the accounting system on the same database. The accounting and accounting are synchronized through a message. The accounting system is used for internal reconciliation. Stopping the accounting system will not affect the business within a controllable time, because it can be restored through reliable messages later. So we made a temporary decision to kill this system and free up 50% of the capacity to get through. This finally made the day pass safely without any danger.

What capabilities does Ant have that it can export to small and medium-sized merchants or external companies?

Currently, the most mature output is to export intelligent customer service technology to our ecological partners to reduce their costs.

In addition, on August 10 this year, Ant Financial's open platform was officially launched, opening up 12 classified capabilities to partners, covering the entire business operation process, including Ant Financial's most basic payment capabilities, exclusive data, security, credit, wealth management, financing capabilities, as well as store opening, marketing, membership, social, collaboration, public services and other capabilities.

2. How is the industry’s top intelligent customer service created?

The business and products provided by Ant’s intelligent service platform department

Financial-grade products have very strict requirements on high availability and fund security. The rapid development of Ant Financial’s financial business has posed great challenges to customer service response and emergency response capabilities.

The core goal of Ant Financial's intelligent service platform is to provide users with a customer service guarantee system with better experience and higher service quality while ensuring rapid business development.

The overall architecture of intelligent services includes seven main parts: My Customer Service (self-service), 95188 (hotline), Service Hall (online), Service Workbench, Operation Workbench, Social Resource Platform, and Public Opinion Monitoring Platform. Currently, the Ant Financial Intelligent Robot (My Customer Service) application has taken over 90% of daily service requests, and the service satisfaction is close to that of manual service.

Compared with other smart customer service technologies on the market, what are the characteristics of Ant Financial's self-developed smart customer service technology?

On November 11, 2015, the entire site answered more than 5 million user questions through self-service, with self-service accounting for 94% and an overall connection rate of over 99%, increasing Ant Financial's customer service efficiency by 20 times.

We are at the absolute leading level in the industry. We can quantify the improvement in efficiency. According to the actual data of manual processing, one service attendant can handle about 100 to 150 user requests for help every day. Based on the estimate of 150 requests, 5 million requests for help require 33,000 service personnel to handle them well. Through our intelligent customer service robots, we can provide users with a service experience that can be accessed anywhere, and the service satisfaction is close to that of manual service.

Ant Financial's business will lead us to provide intelligent customer service for its products. What are the characteristics and difficulties?

To understand this problem, we first need to understand Ant Financial's business background. In everyone's impression, Ant Financial may be equated with Alipay. In fact, Ant Financial currently covers diversified life and financial services from payment, small and micro loans, wealth management, insurance, credit, banking, word of mouth, etc. The diversity and complexity of the service areas have brought great challenges to the intelligence, ease of use and efficiency of customer systems.

We cannot build an independent system for each business unit, so we need to fully consider the platform and scalability of the product architecture from the beginning. Currently, we can quickly launch a new business service system within a week.

Compared with the intelligent customer service technology available on the market, our own intelligent customer service technology has three main differences:

1) Intelligent technology based on large-scale machine learning:

Deep neural networks (DNN) are used for entity extraction, language understanding, and conversation status tracking to better understand user intent. Intelligent IVR uses the world's leading deep learning-based speech recognition and conversation technology to change the menu-based guidance mode of the traditional hotline CC system, allowing users to directly describe the problems they need help with through conversational voice, and understand user semantics for accurate user guidance. User portraits and behavior analysis can accurately analyze the complexity and urgency of user problems based on different user conditions, and intelligently guide users to the most appropriate channels for problem handling. For example, for issues involving security and fraud, it automatically switches to VIP manual service at the first time to provide users with maximum security protection.

2) Complete systematic industry solutions:

The intelligent customer service system is not a simple self-service robot, but a complete engineering system. We provide a complete solution including seven major systems, including my customer service (self-service), 95188 (hotline), service hall (online), service workbench, operation workbench, social resource platform, and public opinion monitoring platform.

3) Enterprise-level SAAS services based on Ant Financial Cloud:

We have taken a solid step towards the goal of empowering the ecosystem with services. We have created a SAAS-based Ant Cloud Customer Service product on the Ant Financial Cloud based on the overall intelligent service solution accumulated from Ant Financial's own business, and empowered external ecosystem partners such as cooperative merchants and institutions.

How to improve the accuracy of questions and answers of intelligent customer service?

1) Extensive application of machine learning algorithms: Deep neural networks (DNNs) are used for entity extraction, language understanding, and dialogue status tracking to better understand user intent. For example, when a user enters the customer service homepage, it will proactively guess the user's problem, predict in advance and guide the user to solve the problem more efficiently, and provide users with caring proactive services.

2) Intelligent IVR technology breakthrough: Intelligent IVR uses the world's leading voice recognition and dialogue technology based on deep learning, changing the menu-based guidance mode of the traditional hotline CC system, allowing users to directly describe the problems they need help with through dialogue voice, and understand user semantics for accurate user guidance. User portraits and behavior analysis can accurately analyze the complexity and urgency of user problems based on different user conditions, and intelligently guide users to the most appropriate channels for problem handling. For example, for issues involving security and fraud, it automatically switches to VIP manual service at the first time, providing users with maximum security protection.

3) Closed-loop operation of knowledge base: We have a relatively complete closed-loop operation system for knowledge base, which can dig out user problems from customer service conversation records based on online user feedback, and produce and optimize a knowledge base that is more in line with the real needs of users.

How to transform the application results of intelligent customer service into universal service capabilities?

This is a very good question. We use productization to support business development. Take the intelligent customer service robot as an example. In the process of developing the customer service robot, we have accumulated a core product, the robot platform, which includes a series of core product modules such as dialogue management, query processing, knowledge base management, user profiling, voice processing, semantic understanding, distribution engine, strategy engine, intelligent IVR, and intelligent quality inspection.

The robot platform itself has good business scalability. In the field of intelligent customer service, we use customer service-related knowledge in domain modeling. In the word-of-mouth business, we only need to prepare domain knowledge related to life services as the robot's knowledge content, and it will be a word-of-mouth life assistant.

Of course, to be a good virtual robot assistant, it is also necessary to combine it with the business content to do more subsequent targeted tuning and customization, which is also very critical.

Excerpts from the Q&A

What is the biggest difficulty in building an intelligent customer service robot system? How would you rate the current intelligent robot customer service platform?

From my understanding, the biggest difficulties of the intelligent customer service robot system include two aspects:

1) Intelligent capabilities : Customer service robots need to approach or even surpass the service level and capabilities of customer service assistants, which is inseparable from the understanding and analysis of data. This process requires comprehensive breakthroughs in data computing capabilities, data application capabilities, and artificial intelligence algorithms. At present, our intelligence level can only reach 70 points, and there is still a lot of room for optimization and improvement.

2) Platform capabilities: From my understanding, a good robot platform must have good scalability. It must be able to face new businesses, quickly output to new business lines, and enable new businesses to innovate quickly. This capability is precisely a full reflection of the capabilities of the middle platform.

This is also a strong foundation for us to empower our ecosystem partners in the future. Only when the platform is well done can it become a good cloud service that can satisfy the service demands of ecosystem partners of different types and sizes. It can also provide our ISVs with customized service products suitable for different enterprises.

How do we get the new business service system online within a week?

This actually has a prerequisite. The underlying platform has indeed achieved decoupling from specific business knowledge, which is what we call a platform-based/product-based approach to making technical products. The prerequisite is that the knowledge of the corresponding business line has been clearly organized. We only need to import the corresponding knowledge base through the platform, and then combine it with general chat, encyclopedia, weather, and other general knowledge capabilities to launch the new business customer service robot within a week.

Of course, as I mentioned just now, if you want to achieve excellent results above 90 points, it requires a long period of knowledge accumulation and continuous optimization, which is a necessary process.

Is the current customer service scenario of Ant Financial different from that of Taobao? The former is more like a salesperson of financial products; the latter is more like what we traditionally understand as Taobao's customer service assistants. If there is a difference, what are the differences between the two in the Alibaba system and where is the synergy?

Your understanding is correct, that is, the groups and business contents they target are different. Ant Financial's intelligent customer service focuses on optimizing financial-related service capabilities, including the possible insurance consultants and financial consultants you mentioned in the future.

On this basis, we hope to accumulate and improve enterprise-level solutions for different financial industries, and empower our ecological partners with overall output. This is also closely integrated with our financial cloud and open platform strategy.

Does your dialogue system directly use deep learning to make decisions, skipping the definition of "dialogue state" in traditional dialogue systems? If so, what is the output space of your dialogue system? (ie action space)

The application of deep learning algorithms in our intelligent customer service has not yet reached the ability to directly skip the definition of dialogue status as you mentioned.

Currently, our robot platform is a relatively complete engineering system.

Different processing strategies will still be integrated. There is a lot of business knowledge that can be "man-defined" and it is still solved by precise templates. You can understand that the dialogue management strategy is a link-like capability, and the strategy can be customized according to specific business scenarios. Different module plug-in access can be customized.

For a newly added field, such as "Ask about the [rate of return] of a [specific insurance product]", if this field has not been done before, how quickly can your current system support it? Especially when such requirements are more complicated and large in volume, without a unified framework, how can your system cope with it?

This is challenging. What we are trying to do is the ambiguity you mentioned, or the ability to quickly learn new business knowledge. Our artificial intelligence team, that is, Dr. Qi Yuan's team, is doing research and application in this area. If possible, we can invite Dr. Qi Yuan to have a more in-depth exchange with you next time.