Shulou(Shulou.com)06/02 Report--

In the field of artificial intelligence, intelligent customer service is an application practice with relatively easy landing and mature technology. This paper takes intelligent customer service as the object, and combs its development process, system construction and market promotion. Enjoy~

Google demonstrated the conversational robot Duplex at the 2018 IPUBO developer conference.

Duplex accomplishes two tasks:

The first task is to book a haircut service; the second task is to book a telephone reception for dinner.

In fact, Duplex plays the role of smart customer service.

In the field of artificial intelligence, intelligent customer service should be relatively easy to land, and the technology is relatively mature, this is because the scene path in the field of customer service has relatively clear characteristics, which determines that artificial intelligence based on full data for high concurrent demand processing will have a bright future in the field of customer service.

At present, based on the leading artificial intelligence technologies such as big data, cloud computing and deep learning, intelligent customer service has been able to achieve a series of complex operations, such as independent question and answer, business processing, fault diagnosis, and so on, to achieve most of the response requirements in the customer service industry. solve user problems quickly and efficiently.

According to the China Intelligent customer Service Industry Research report released in May 2018, there are about 5 million full-time customer service in China, with an average annual salary of 60, 000, plus hardware and infrastructure, with an overall size of about 400 billion yuan.

Such a huge market, of course, will make many enterprises flock to intelligent customer service. But why hasn't a unicorn company emerged yet?

Although this is the easiest and relatively mature project in artificial intelligence, how much does it cost for relevant enterprises to develop and build an artificial intelligence customer service system?

Whether an enterprise should build a set of intelligent customer service system by itself, or find a suitable intelligent customer service platform manufacturer, stand on the shoulder of the "giant" and use the ability they give to build its own intelligence.

Able to provide customer service solutions.

Let's have a good talk today.

I. the development of customer service system

Chinese customer service software market has roughly experienced three stages of development: traditional call center software, PC web page online customer service + traditional customer service software, cloud customer service + customer service robot intelligent customer service stage.

Before 2000, the Internet was not yet popular, and customer service was mainly based on telephone communication. From 2000 to 2010, thanks to the integration of computer technology, computer telephone integration technology (CTI), network technology, multimedia machine technology and enterprise information applications such as CRM, BI, ERP, OA, etc., customer service system jumped out of a single telephone communication and appeared a variety of customer service channels such as web page and online customer service. In the past decade, the development of mobile Internet, cloud computing, big data and AI technology has brought the traditional call center and customer service software into the SaaS and intelligent era. On the one hand, the brand-new SaaS model greatly reduces the cost of building customer service centers, and the SaaS model is becoming more and more popular. early manufacturers of call center hardware equipment have been extended to the middle and lower reaches to provide local customer service center solutions for large customers such as foreign enterprises and state-owned enterprises.

From the current composition of the customer service industry chain, the upstream infrastructure link has been mature, a small number of giants monopolize the market. In the future, they will continue to extend downstream to build an enterprise service ecology.

Among the mid-stream customer service product providers, cloud customer service manufacturers have come to the fore after several years of competition, but they have not yet grown into giants, and the competition is still fierce. The product features are more abundant, and the application scene extends from customer service to sales, marketing and other links. On the other hand, the customer service robot greatly improves the efficiency of artificial customer service by assisting manual workers and answering simple and repetitive questions. At the same time, AI is also changing the interaction mode of enterprise customer service from various aspects, accelerating the intelligent upgrade of online and offline customer service.

Second, the construction of intelligent customer service system

The intelligent customer service system is mainly based on natural language processing, large-scale machine learning and deep learning technology. it uses massive data to build a dialogue model, combines multi-round dialogue and real-time feedback autonomous learning, and accurately identifies user intentions. support text, voice, pictures and other rich media interaction, can achieve semantic analysis and multi-form dialogue.

Task dialogue service:

Customized service, through multiple rounds of interaction with users, to achieve express inquiry, food ordering, doctor pre-diagnosis and other service functions.

Business consulting services:

Through the QA knowledge base, quickly reply to the user question consultation service. Answers to frequently asked questions.

two。 Technical Framework of Intelligent customer Service system

(1) Intelligent customer service system based on knowledge base answer.

Intelligent customer Service system based on knowledge Base answer

Implemented using a retrieval or classification model.

The process of retrieving answers is:

First, deal with the user's input problems, such as word segmentation, keyword extraction, synonym expansion, sentence vector calculation, etc., and then do retrieval matching in the knowledge base based on the processing results, such as matching a problem set using BM25, TF-IDF or vector similarity, which is similar to the recall process in the recommendation system. Since we are a question and answer system, we finally return an answer to the user directly, so we need to pick out the most similar question from the question set. Here, we will reorder the question set, for example, using rules, machine learning or deep learning model to sort each question, and each question will be given a score. Finally, top1 will be selected, and the corresponding answer to this question will be returned to the user. This completes a conversation process.

In practical application, we will also set a threshold to ensure the accuracy of the answer. if the score of each question is lower than the threshold, several questions in the head will be returned to the user in the form of a list, and the end user can choose the questions he wants to ask. and get specific answers.

(2) Multi-round dialogue system based on slot filling

Building a slot-based dialogue system is a relatively professional and complex process, which is usually divided into three main stages. The first is requirements analysis, then the use of the platform to build BOT, and finally continuous optimization.

To understand the system, let's first familiarize ourselves with the definitions of several nouns:

1) intention

Intention refers to the main request or action made by the user in the voice interaction.

Example of intent:

Positive intention: yes; right; correct; Ok; negative intention: no; wrong; wrong; NO; cancel intention: exit; stop; close; end

2) skills

Skill is an application that meets the specific needs of users. For example, when users say, "query where my shampoo is delivered", they will enter the skills of express query.

3) question and answer skills

Through the configuration of Q (user question) and A (robot answer), a simple dialogue between the user and the robot can be realized.

Task-based skills: on the basis of question-and-answer skills, add slots, API (interface) calls and other advanced functions, which can be configured to achieve user query information, question search or other functions.

4) Dictionary

The content in which a keyword may change, such as a time dictionary, a location dictionary.

Semantic slot: semantic slot is a key word contained in the user's statement, which can help the system accurately identify the intention, such as the constellation semantic slot contains the name of 12 constellations. Semantic slots and dictionaries are usually used at the same time, and semantic slots are usually used to refer to dictionaries. A semantic slot can bind multiple dictionaries at the same time, and a dictionary can also be associated with different semantic slots.

5) follow-up

When the semantic slot value is not provided in the user question method, the robot will automatically initiate questioning against it.

For example, the user asks: what's the weather like? We can't get the semantic slot value of the place where the weather is queried, so we need the robot to ask, where do you want to get the weather information? Follow-up questions are generally set up with multiple questions, which are asked at random.

In the domestic open bot system, the dialogue open platform between Baidu UNIT and Wechat is the technical framework of the application.

In a natural language dialogue system, the core task of understanding is the analysis of intention and the identification of word slot.

For example: book the train from Beijing to Shijiazhuang at 8 o'clock tomorrow morning. In this example, for a sentence expressed by the user, its intention is to book a train ticket, including the place of departure, destination and time. When there are multiple trains at this time, you need to ask the user which one to order.

Take Baidu UNIT platform as an example to build a process of intelligent reply for ticket purchase.

Demand analysis: booking a train ticket needs to know the time, place of departure and destination to create a new BOT, named: train ticket new dialogue intention: name booking add word slot: departure time, select system word slot dictionary, select and then select system dictionary sys_time (time), departure word slot, destination word slot, these two can choose system dictionary, these are required items. Set the attribute of association between word slot and intention, where the departure time of the train ticket is the necessary key information in the booking, so select the required information. Clarification means that bot actively asks the user to clarify when there is a lack of departure time in the statement expressing the booking demand. You can also set the number of rounds for users to clarify and then abandon the request for clarification. The default is 3 times. Set up the BOT response, which is the feedback given to the user when the BOT identifies the user's intention and all required word slot values. For booking reply, it is generally interfaced with API API to realize automatic generation.

Of course, this is only a scene in the train ticket, in which there are functions such as refund, ticket change, inquiry and so on. These are what we need to determine in the needs carding.

3. How to judge the quality of an intelligent customer service system

(1) Evaluation based on manual labeling

In the system based on the question and answer knowledge base, the answer ability is limited by the richness of the knowledge base, that is to say, the higher the coverage of the knowledge base to user questions, the higher the accuracy.

Therefore, it is not able to answer all the users' questions, and the best state of the system is to reject all the answers that can be answered accurately and those that cannot be answered, that is, refuse to answer.

Therefore, the evaluation indicators here include problem resolution rate, rejection rate, recall rate and accuracy and so on. Our goal is to make the result rate of the system infinitely close to the real result rate of the data, and the recall rate and accuracy rate as high as possible.

Recall rate = number of questions that the robot can answer / total number of questions accuracy = number of questions correctly answered by the robot / total number of questions solved by the robot = number of questions successfully solved by the robot / total rejection rate = number of unanswered questions / number of user questions

A small data set is sampled from the daily full data set to ensure that the data distribution of the small data set accords with the full data set as far as possible, and then the annotation team marks the data set and marks the actual answer to each question. Generally, after the labeling is completed, there is also a link of quality inspection to ensure that the labeling results are as accurate as possible, so as to generate a standard evaluation set of daily data.

Based on the standard evaluation set, we will evaluate the quality of the system, and each iteration of the new model will use the standard evaluation set to evaluate the new model, only when the new model reaches a certain index can it be online.

(2) Evaluation based on user feedback

Manual evaluation can evaluate the accuracy of intelligent customer service system, but whether the answer is reasonable and whether it can solve problems for users needs feedback and evaluation by users. the ultimate goal of the whole intelligent customer service system is to help users solve problems.

We will design the evaluation functions of intelligent customer service and online customer service on the product, for example, let users evaluate each answer or a conversation of intelligent customer service, and send an evaluation card to users to evaluate their satisfaction after chatting with human customer service. as shown in the following figure.

Finally, we will count the proportion of participation, satisfaction and other indicators, these indicators can really reflect the quality of the intelligent customer service system. In practice, the proportion of user participation is often low, we will use a variety of methods to stimulate user evaluation.

Third, those problems encountered by intelligent customer service 1. General intelligent customer service system or vertical industry intelligent customer service system

The intelligent customer service system is 2B, and the universal intelligent customer service system means that the market is bigger and there are more users. On the other hand, there are much fewer customer service system users in the vertical field.

Take the insurance industry as an example, there are more than 100 insurance companies across the country. And to do intelligent customer service systems in the vertical field, the AI team must fully understand the industry. Understanding business needs and understanding business processes also requires cross-departmental communication.

Do the intelligent customer service system in the vertical field, often fall into one or two big projects, constantly meet the personalized needs of users. In the end, the system is very "customized" and the market is very small. After doing a few projects, you will encounter a transparent ceiling.

However, the market for general-purpose intelligent customer service systems is very large, but compared with the teams that do intelligent customer service systems in the vertical field, there is no advantage, there is little gap in technological advantages at the present stage, and small companies can customize users. But the universal system can not, and eventually become a very large market, but it has been eroded by small companies that do intelligent customer service systems in the vertical field one by one.

What are we going to do?

At the beginning of the Internet, portals took the lead in serving the needs of most people. Next, the official Wechat account can be subscribed to, and everyone's reading content is different. This is a customized version of the information platform. From the user's point of view, customization is the direction of evolution, and eventually universal customer service will be replaced by vertical industry intelligent customer service.

two。 SAAS service or privatized deployment

Large enterprises in traditional industries, such as banking, insurance, securities and real estate, often have a strong demand for customer service and a strong willingness to introduce intelligent customer service systems, but at the same time, they also have high requirements for their own data security, so they will only agree to localized solutions.

When such major customers do localized deployment solutions, they can only adopt the project-based business model and charge a fee for a project. The advantage is that a project can receive tens to millions of yuan of income, and it can be profitable at the initial stage of starting a business; the disadvantage is that customers for privatization deployment need a lot of customized needs, which will take up a lot of labor costs and are difficult to replicate on a large scale. in the long run, the room for growth is limited.

What are we going to do?

From the perspective of data security alone, it will be solved with the development of technology. At the beginning of mobile payment, people were afraid that binding their own bank cards would be stolen. Will hackers hack into my Alipay. Now it seems that the worry is unfounded. With enough investment, there will be enough funds to support technology development, there are more users of SAAS services, technical loopholes can be found more easily, and the security of the system will evolve faster. Privatization deployment is not a good option.

3. Serve large customers or small and medium-sized customers

When choosing target customers at the beginning of a business, all smart customer service startups are faced with a choice: whether to focus on large corporate customers, or to cut into the SME market in the first place?

Main cut small and medium-sized enterprise customers can use standardized SaaS products to meet their needs, not only the mode of light occupation and low labor cost can achieve large-scale replication, but also can obtain continuous income by the way of annual renewal, and can constantly get data circulation feedback to establish technical barriers.

But the disadvantage is that it is difficult to get customers in the early stage, a lot of market education work needs to be done, and the mortality rate of small and medium-sized enterprises is high, the overall renewal rate is difficult to guarantee, and it is difficult to make a profit at the initial stage of starting a business.

However, if you focus on major customers, some customized needs are difficult to meet, and the process of major customers is relatively long, and service providers with long-term service generally have high requirements for product maturity, so it is difficult for startups to enter. Positioned to serve several major customers, it is relatively risky for startups.

What can we do?

To do a vertical SAAS system, we need more users to use it in order to iterate the system faster. With only one or two big customers, it is difficult to put forward constructive suggestions for improvement, so as to be small and medium-sized customers, find the first batch of users as soon as possible, run the system and then constantly optimize and iterate.

3. Sales difficulties of intelligent customer service

Everyone is saying that there are many pain points in the traditional customer service industry, and intelligent customer service can solve these pain points very well. For example:

(1) High labor cost

When the demographic dividend disappears, the employment cost of the employer will be higher and higher.

Is this the real demand? First of all, customer service is not the core department of an enterprise, and most enterprises do not attach great importance to the customer service department. In small and medium-sized enterprises, there are not too many customer service personnel, and the real labor cost savings are not high, so the motivation for enterprise replacement is not great. In large enterprises, human cost is indeed a large cost department, but it is also based on this, large enterprises have enough expenditure to do their own intelligent customer service system. Because their input-output ratio is appropriate. Like companies with large customer service departments such as Didi, they prefer to do it on their own.

(2) decision paradox

Intelligent customer service systems solve what human customer service does, and when their jobs are replaced, it means layoffs in the department.

Of course, this is a good way to cut costs for enterprises, but it is not so good for the leaders of the customer service department, who say that the reduction means that they have a lower weight in the enterprise.

Although this is the general trend in the long run, now the sales process is basically a top-to-bottom sales process, rather than the urgent needs put forward by the department, and the department personnel continue to follow up.

Summary

There is nothing new in the sun, and big companies use the underlying technical framework to build their own intelligent customer service systems. It may be a trend to ensure the security of data as well as to control costs. For some SAAS intelligent customer service systems, when technology does not form an oligopoly, product promotion and service capabilities will become particularly important.

Are there any barriers for smart customer service companies? What are the barriers to smart customer service companies?

The usage habit of customer service system, the accumulation of data, and the improvement of knowledge base are the industry barriers of intelligent customer service system. The cost of switching intelligent customer service system is too high for users to replace it.

So expand your users as soon as possible, this is the barrier of intelligent customer service companies. Only smart customer service business growth in the future will be very limited, to find their own second growth curve is the key to determine how far the intelligent customer service company will go.

Http://www.woshipm.com/ai/2840619.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.