In recent years, I have been engaged in intelligent customer service. In the process, I also participated in the functional design of some IM products. Here I will briefly talk about my thoughts on IM product design.

After chatting about IM, customer service IM and call center products in the previous few articles, today I will talk about the intelligent customer service question-and-answer robot which has been quite hot in recent years.

The background,

On the official website of domestic customer service system manufacturers, there will generally be such product solutions: omni-channel intelligent customer service products or platforms. Today, I would like to talk about a few key words here: omni-channel, intelligence, platform.

1.1 all channels

The traditional customer service channel is the telephone. Later, with the development of the Internet, mail, web, various social products, various apps have emerged, and the channels for customers to respond to problems have also become more and more.

With more channels, the customer service system needs to support more functions. Each channel has its own characteristics, and the agent workbench and question answering robot need to provide unified interaction and the same service. Therefore, customer service products need to provide more intermediate matches in the channel access layer.

For example, how to support multiple customer service by mail server, how to support rich text by wechat public number, how to achieve complex information collection by phone, and how to mix online customer service and telephone customer service.

So omnichannel means two things:

  1. Ability to access customer consultation through multiple channels
  2. Ability to integrate multiple channels and provide unified interaction and customer service.

Second, there are many challenges to be faced in product realization, the most difficult one being the unification of telephone and text.

1.2 smart

General customer consultation is the source and c-terminal users and customers, the characteristics of its consultation are:

  1. A large amount, I think this is also the Internet has been rumored Tencent no artificial customer service reasons.
  2. The similarity of problems is relatively concentrated, 80% of customers’ consultation problems are concentrated in 20% of the problems, in line with the 80/20 principle.

Based on this background, there is a need for intelligent customer service.

The intelligence of intelligent customer service is embodied in intelligent reply, which is the ability to automatically reply to customers’ questions by using AI technology and NLP algorithm.

Whether through phone or online channels, automatic response can be achieved through channel adaptation and transformation, which greatly reduces the workload of customer service and allows customer service to focus more on complex problems that robots cannot solve.

1.3 platform

Technology and software companies like to talk about platforms, and customer service product manufacturers are no exception. They like to use platforms to explain the versatility, universality and power of their products.

That intelligent customer service products, how are the characteristics of the platform reflected?

My personal understanding is that the integration degree is high. It integrates all subsystems and sub-products of the customer service system into one set of products to provide one-stop systems and services for enterprises, saving the pain points of integrating multiple products and systems.

For example, the intelligent customer service platform has achieved all-channel access capability, so it needs to provide:

  1. Telephone customer service: switch or softswitch
  2. Online customer service: access to multiple social and IM platforms
  3. Intelligent customer service: text robots and telephone robots
  4. AI platform: training and updating of models.

Here any of the sub-products and systems of the workload is very large, can be integrated together, its difficulty is not small, so the real need for customer service system for the enterprise, is also a great attraction.

In addition, enterprises that generally need customer service have already had their own call centers (Cisco, Genesys, Avaya, Huawei, China Telecom, China Unicom and China Mobile) for a long time. If intelligent transformation is needed, the migration and adaptation of the original system are also points to be considered.

Second, intelligent customer service

Because of the different ways of interaction, but also brought about different technical implementation, customer service requirements are also many different, from the previous several introductions, we can know that customer service system according to the way of communication, divided into telephone customer service (voice) and online customer service (text).

2.1 Intelligent Online Customer Service (text)

With advances in AI and NLP technology, intelligent customer service products have become more common these days. We see it everywhere: navigation, shopping, smart homes, and so on.

From the background of the emergence of intelligent customer service, it can be seen that it is a complement to traditional customer service products. It is a tool to improve customer service efficiency and reduce customer service pressure through AI technology and NLP capability.

The development of online customer service came a little earlier than the development of robophone, and it depends on the maturity of the technology.

The product structure of an intelligent online customer service is as follows:

2.1.2 Robot Configuration

The configuration of robots is divided into several aspects:

  • Enterprise information configuration: this part is mainly used in the web chat window

  • Basic configuration: includes core configuration items and scenario-specific configurations

  • Channel interconnection configuration: Complete the connection with the service system

2.1.2.1 Enterprise Information Management

Enterprise information management mainly sets the enterprise name, Logo, email, official website address, and enterprise profile.

  • The enterprise name and Logo affect the Logo and name in the upper left corner of the management background.
  • The enterprise name and Logo, official website address, and enterprise profile affect the header information in the chat window.
2.1.2.2 Robot Configuration

The basic configuration is mainly used to manage various parameters and scene of the robot. All configuration points and methods are given default values. During the initialization phase, the robot can be used without setting or modifying.

  • Basic setup

These include:

  1. Robot name: Mainly used as the default name of the robot in the channel
  2. Robot avatar: Mainly used as the default name of the robot in the channel
  3. Robot speech: The language here is mainly used to build the robot question and answer engine can support the speech, a service instance intelligently support a speech
  4. Use word vector: after this function is enabled, word vector will be used to refine the answers to be arranged based on retrieval and similarity calculation to improve the accuracy of question and answer.
  5. Enable the greetings database: after enabling, the robot will use the greetings corpus as the bottom of the question and answer process, such as: Hello and other questions
  6. Direct response confidence: the value set here affects the score judgment of the direct response of the robot, and the answer whose similarity is greater than this value will be directly replied.
  7. Confidence of unknown reply: The value set here affects the score judgment of the unknown reply of the robot. When the similarity is less than this value, the robot will give a reply. Between the unknown reply and the direct reply, the robot will give a rhetorical guidance.
  8. Visitor offline time: refers to the time when the user does not interact with the robot, the session will be automatically offline
  9. Enable answer satisfaction evaluation: After this function is enabled, when the robot replies to questions asked by visitors, it will guide users to evaluate whether they are satisfied with the answer
  10. Session satisfaction evaluation: After this function is enabled, the session satisfaction evaluation invitation will be launched at the end of the session, asking users to comment on the current session.
  • Words set operation

The setting of speech technique is mainly to help the robot to complete the reply in various fixed scenes, such as how to reply when the problem cannot be recognized.

Multiple fixed speech techniques can be set for each one. When the robot uses speech techniques, it can choose random selection or use all of them according to the specific scene

All robot related words are as follows:

  1. I guess you want to ask these questions:
  2. Oh, this is too difficult. I haven’t studied it yet. Can you change the question
  3. Sensitive language: Your words contain sensitive information, please be careful with your words
  4. I guess you want to ask:
  5. Thank you for your encouragement. I will try my best to answer the questions better
  6. Answer satisfaction point step on words art: woo, failed to solve your problem, I am very sad, woo
  7. Please point out what I did wrong so that I can correct it in time
  8. Answer satisfaction unsatisfaction reason collecting follow-up tips: Woo, I will admit the mistake to change, look forward to my change oh
  9. Give me a score for my service
  10. Thank you for using it. I will redouble my efforts to provide better service
  11. Conversation satisfaction evaluation Other Reasons Collection Invitation Conversation: Anything else you want to say to me
  12. Visitor offline prompt: I have been offline, thank you for your company, if you need further consultation, please resend the question
  13. Answer evaluation is not satisfied with the reason: answer is not asked, unresolved problems
  14. Cause of unsatisfied session evaluation: unsolved problems, inaccurate matching, and unintelligent
  15. To better serve you, please select a skill group:
  16. Transfer to manual initiating prompt: You are being transferred to manual service
  17. You can continue to ask me questions
2.1.2.3 Channel Connection

Channel management is mainly used to manage the connection between business system and intelligent customer service Q&A interface, such as how to manage the web chat window link embedded in business system, server configuration parameters of social platform, etc.

2.1.2 Knowledge management

After completing the robot configuration, we moved on to the knowledge management section. Knowledge management is divided into three parts: semantics, corpus and labels

2.1.2.1 Semantic Management

Semantic management is mainly responsible for managing all kinds of professional words, sensitive words and stop words needed in the robot question answering word segmentation engine.

  • Professional word: refers to the specific business area of the established vocabulary, such as foreign exchange, order, delivery, express delivery, such as robot built-in participle engine can only handle general common words in the field of library, for a specific business area of thesaurus, model absorb less, through professional word, word segmentation can be effectively increased the accuracy in engine, raise the accuracy of the answers.
  • Sensitive words: words that you do not want customer service to mention during the question and answer process, such as abusive words, pornography, violence and so on.
  • Stop words: This part is to ensure that the robot pays more attention to the core thesaurus during the question and answer process, so some common auxiliary words and prepositions are used as stop words, and words that often appear in specific industries and have little effect on semantic understanding can also be added here.

In addition to word types, there are two kinds of relationships:

  • Synonyms: The record here is that two words have the same meaning, in the robot understanding and recognition, can be equivalent replacement.
  • Enumerator words: here is to record two words with upper and lower relationship, such as plant and fruit, fruit and apple. When encountering this word, the robot can trace up and drill down to improve the ability of semantic generalization.

Management function in the management background -> semantic management -> thesaurus management, any modifications made on the interface, 15 seconds after the robot question and answer interface can take effect.

2.2.1.2 Knowledge management

Knowledge management mainly manages the knowledge base required by robots, which is divided into two parts: business knowledge and communication knowledge. It provides functions of adding, modifying and deleting knowledge.

A very important feature here is sandbox management. Considering that knowledge is very core to the robot, knowledge management is provided. By default, sandbox mechanism, online knowledge and sandbox knowledge are provided.

  1. By default, the added and edited knowledge is in the sandbox, and the status is to take effect. By using the online function, the status is to take effect, and then the robot can use it.

  2. Knowledge-sensitive operations provide an audit mechanism. For users who are not administrators, when they initiate online operations after editing knowledge, they need to be reviewed by the administrator before they become online. For offline and delete online knowledge, they also need to be reviewed before it takes effect.

  • Business knowledge

Business knowledge is mainly THE FAQ accumulated in a specific industry, which is generally related to business, such as how to exchange points.

Knowledge consists of standard question, similar question, answer, knowledge label and answer label.

  • Standard and similar questions are used to match the user’s questions
  • The answer is the reply to the chat screen after the match is successful.
  • Knowledge tag and answer tag are used to achieve knowledge isolation, for different channels, different users, can be asked about different knowledge and different answers. The tag data here comes from tag management

Similarity query management allows you to add, modify, and delete similarity queries. The changes take effect only after they take effect online.

  • An intellectual

The difference between greeting knowledge and business knowledge is that greeting knowledge does not involve business, knowledge content is about all issues outside the business, such as the image of robots, gossip and so on.

The function of knowledge management is the same as that of business knowledge management. There is one more function, which is the built-in corpus of robot image. In the initialization stage, the built-in greeting corpus can be loaded to quickly build the knowledge related to robot image.

In addition, the robot can use the greeting knowledge, also need to enable the greeting library in the robot Settings.

2.3.3 Label Management

The label management function is mainly used to deal with knowledge isolation, which can be implemented in the following scenarios:

  1. Different people can question and answer different knowledge: through the knowledge tag.
  2. Different channels can respond to different business scenarios: through knowledge tagging
  3. Different people ask the same knowledge, can give different answers: complete through the answer tag.
2.3.3.1 Label Management

Labels are managed in a tree structure. Labels at different levels have the capability of upward compatibility. For example, labels in Jiangsu can cover labels in Nanjing, thus achieving upward compatibility and downward refinement of binding knowledge.

A label has multiple conditions, which can be and (both satisfy) and or (any one satisfies).

2.3.3.2 Condition Management

The condition here is the filtering rule corresponding to the indicator. The information carried by the user identity bound to the label is set to filter the knowledge bound to the label.

Currently, you can bind user identity information including user name, gender, region, and channel.

The filtering logic is as follows: the personal information carried by the user when entering the chat window and the tag of the knowledge and answer obtained by the robot after asking questions are compared with the user’s identity information based on tag association conditions. If the information is the same, the user can ask the knowledge and answer. If the information is not the same, the unknown reply will be given.

2.1.3 Question and answer service

After completing a series of configuration and knowledge initialization of the robot, we can question and answer in the QUESTION and answer interface.

The basic functions are as follows:

  1. Support small talk and business q&A
  2. Support the guidance of rhetorical questions
  3. Support message and satisfaction evaluation
  • Ask guide

For visitors’ questions, if the robot finds that there are multiple possible knowledge after recognition, it will give rhetorical guidance.

For example, if you ask an employee, the real intention of the visitor cannot be clearly defined. At this point, the robot will provide a list of employee-related questions for the second confirmation by the visitor. After clicking on the specific question, the visitor will give an accurate answer.

  • Unknown to reply

When the robot fails to recognize the user’s problem, it will give an unknown reply. The judgment is based on the confidence of the unknown reply in the configuration of the robot. The given reply is the position of the reply in the configuration of the phone.

  • Leave a message

When the robot wants to transfer to a human, but the human customer service is not online, the problem can be collected through the message function.

The collected messages can be viewed and processed in the background message management. When leaving a message, the contact information is provided. At this time, the customer service can inform the result after processing through the contact information.

  • Session satisfaction evaluation

When the visitor does not interact with the robot for a long time and reaches the timeout period, the robot will push the session satisfaction invitation review.

2.1.4 Access to human customer service

In order to provide better customer service, the human-in-the-loop mechanism is generally guaranteed behind intelligent customer service to provide manual service. The most important thing here is the transition to manual strategy.

  1. Configure the channel to manual policy
  2. 2.5.1 Skill Group Management

The skill group is a virtual group established to distinguish the skills and experience of customer service. Normal skill groups can be divided into pre-sale, in-sale, after-sale and other types. If the business of an enterprise is more complex, the difference can be more detailed.

After completing the configuration and maintenance related to customer service, the next step is to configure the human-to-human strategy corresponding to the robot in a specific channel.

To manual policy classification to manual mode and binding skill group:

2.5.5.1 Switch to Manual mode

There are four ways to transfer labor:

  1. Robot first: After entering the chat window, there will be robot service first, and visitors can initiate manual transfer
  2. Manual first: After entering the chat window, it will automatically transfer to manual. If manual is busy or offline, it will continue to be served by the robot
  3. Robot only: After entering the chat window, the service can only be provided by the robot
  4. Customer service only: after entering the chat window, it will automatically transfer to the manual. If the manual is not online, the session ends and the robot does not provide service
2.5.5.2 Switching to a Manual Policy

After the manual mode is selected, you need to configure the passive manual mode. Generally, the following types of policies are provided:

  1. Answer point tap trigger: When the answer point tap reaches the configured threshold, the system automatically switches to manual
  2. Triggering unknown problems: When the number of triggering unknown replies reaches the configured threshold, the system automatically switches to manual.
  3. Trigger of dissatisfaction: after the dissatisfaction is identified in the question, it will be automatically transferred to manual
  4. Turn to manual questions: Set which questions will automatically turn to manual
  5. Go to Manual button: A go to manual button is available on the interface.
2.5.5.3 Skill Groups corresponding to channels

You need to configure the skill groups that can be accessed by each channel.

  • If multiple skill groups are bound, you need to select a skill group when the visitor of this channel is transferred to an employee
  • If there is only one skill group, you go to the customer service under this skill group by default
  • If no skill group is configured, the system automatically assigns skills.

2.2 Smart Phone Customer Service (Voice)

Smart phone, its core essence is still through the question answering robot to complete the automatic reply of business questions, but because of its different interaction mode, the product depends on some different technologies.

2.5 ASR and TTS

First of all, smart phones need to rely on the technology to convert text and voice. The two technologies are generally used together, namely, language to text and text to voice. The voice file must meet the 8K sampling frequency to communicate with the SIP server.

A few years ago, Iflytek of The University of Science and Technology has a very high voice recognition ability, which can recognize a variety of dialects. The voice recognition ability here is ARS, which converts speech into text. After turning into text, we can realize question and answer in the category of text robot. After receiving the answer through the text robot, we broadcast the text to the customer through the text to speech technology, so as to realize the automatic question and answer on the phone.

2.5 Telephone Robot

Phone robot is a blend of q&a robot IVR system, through the SIP protocol communicate with call center, robot through call center to provide voice information to phone, phone robot through the ASR transcribed into words, again through the semantic analytical intention recognition engine, based on several rounds of session engine and knowledge base, to give reply, is converted to voice through the TTS will reply, The call is sent back to the call center to complete a round of interaction.

In this process, there are many exceptions to handle:

  • interrupt
  • Environmental sound
  • Insert asked
  • dialect
  • Accent.
  • Differences in language expression
  • .

All these problems need to be well dealt with in conversation management to ensure the service quality and experience of telephone robot.

To solve these problems, many capabilities need to be integrated, such as text robot, session engine, ASR and TTS. These capabilities are generally in a Pip line, so their accuracy is not simple accumulation, but multiplication, so the error will be amplified.

2.6 Outbound Robot

This function is more applied in marketing, notification, information transmission and other scenes, such as express delivery, overdue reminder, epidemic prevention notification, new product release notification, etc.

This reference is made to the smart phone customer service architecture of the call center vendor, which is the same as that of most vendors.

Question answering robot

Whether it is intelligent online customer service, or smart phone customer service, its internal core or question answering robot, question answering robot realizes natural language as input, get natural language output, complete the answer to the question.

3.1 Stage Division

Its development has gone through several stages, broadly speaking, search engine is also a rudimentary version of question answering robot.

The level of question-answering robot is defined by Gavin, the developer of star Intelligent question-answering robot

  • Level 1 is based on the command interaction mode. 10 years ago, for example, terminal interaction on the command line required precise commands to obtain accurate answers.

  • Level 2 question answering robot is based on FAQs retrieval and similarity calculation, which is also the realization form of most question answering robot manufacturers at present

  • Level 3 Contextual understanding: In the question and answer process, the user’s contextual information can be understood and automatically associated recognition. Few question and answer robot manufacturers have reached this stage.

  • Level 4 Advisory robots, which could be available in the next two to five years, combine recommendations and knowledge graphs to adjust the robot’s state in real time based on the content of the communication

  • Level 5 Adaptive robots can be adapted in any domain, providing versatility and limited domain q&A.

3.2 Problems to be solved

So what kind of problems does a question-answering robot have to solve?

Question answering robots are designed to:

Start with the question and get the answer to the question.

Therefore, it can be divided into four stages: problem acquisition, problem analysis, knowledge matching and answer return.

3.2.1 Get the answer

Level of problems, mainly dealing with access to a variety of channels, such as WeChat, small programs, such as nailing, identity authentication, third-party business system integration, which is embedded in a particular business system, such as OA, various kinds of APP, interact, there may be text, may also be speech, need to deal with the TTS and ASR, There may also be rich media image processing that needs to provide OCR and image recognition.

  • Channel access
  • The identity authentication
  • Third party Integration
  • ASR
  • OCR
  • TTS
  • The picture video

3.2.2 Problem analysis

Our everyday dialogue habits of people, typing habits, identified in the problem analysis phase, we need to handle good word segmentation and generalization, the need to consider the context of dialogue, well completion, need to consider the input error, peculiar to Chinese pinyin fault-tolerant, glyph fault-tolerant, specific intent discriminant of dialogue, think about what the user wants to express, Which information needs to be focused on and recorded? Based on the tone of civilized communication, we also need to deal with the detection of forbidden words and sensitive words, etc.

  • Semantic word segmentation
  • Sensitive word detection
  • Refer to eliminate
  • Intention recognition
  • Entity Identification (NER)
  • Completion to rewrite
  • Fault-tolerant processing

3.2.3 Knowledge matching

After problem analysis, we get the word segmentation, generalization, normalization, correct expression, clear intention of the question, this time is the time for each robot engine, facing a variety of different scenarios, knowledge performance types, robots need to meet the FAQ engine based on information retrieval. Based on task of several rounds of dialogue, form the robot based on tabular data, based on the document reading comprehension of the robot, based on knowledge atlas of the robot, and at the same time in order to keep the robot’s interesting and anthropomorphic, we also need to various types of preamble, check the weather, tell jokes and banter, for example, in the field of a particular business, we also need to integrate business systems, Help users to implement personal personalized query services, such as check points, check balance, check bills and so on.

  • FAQ-Bot
  • Task-Bot
  • Table – Bot (Text2SQL)
  • ChitChat-Bot
  • MRC-Bot
  • KG-Bot
  • Business Integration (Action)

3.2.4 Return the answer

Through multiple engines, we get the final for the answer, may be several rounds, may be a FAQ, may also be a form or a preamble, with multiple options, we need to decide how to push the answer, at this moment need to combine channels, visitor status, knowledge domain filtering on answers, assembly, in order to achieve a particular end, For example, text on wechat terminal, voice and text on physical terminal, etc., will finally present appropriate answers to users in appropriate ways.

  • Channel segregation
  • User roles
  • Multiterminal adaptation
  • The answer to assemble
  • Field of isolation

The above is my personal understanding of intelligent customer service products.

At this point, the entire IM product understanding and thinking on the introduction of the end, I hope my understanding can give you to inspire the role.