Design method of intelligent customer service dialogue scene

1. Definition and classification of scenarios in intelligent customer service

Users need to have many conversations with robots to solve a problem, which is the dialogue scene in intelligent customer service. Generally, it can be divided into single scene dialogue and multi-scene dialogue.

1. Single scene dialogue

Definition: users and robots only talk about one topic, such as "checking the weather"

? Example 1? Scene check weather

User: What will the weather be like tomorrow?

Bot: MINUS 10 degrees. You can wear more.

User: What about Shanghai?

Bot:: negative 1 degree.

2. Multi-scene dialogue

Definition: The user contacted the robot and had a conversation on several topics such as "order inquiry+return".

Example 2? Scenario Query Order+Return

? User: Look where my shopping went.

? Bot:？ Issue orders to users

? User: Well, look at the second one.

? Bot:？ Publish order details

? User: It hasn't been delivered yet. Please return it for me. Judging from the user's words, the scene of the conversation changed from "asking about the order status" to "returning goods". Returning goods and inquiring orders are two different topics, which can exist independently or be combined into one.

.

2. Scene design

Design a dialogue scene, which mainly includes two aspects: starting condition arrangement and process arrangement.

Little knowledge: suppose "inquiring about the weather in Shanghai tomorrow" is the first sentence spoken by the user, then it is the starting condition in a scene, and in this sentence, it consists of two element words "tomorrow" and "Shanghai", which represent the time and place respectively. It can be inferred that the main structure of this initial condition is time+place.

1. Arrangement of initial conditions

When the user interacts with the robot for the first time in the dialogue scene, what he says is the starting condition. However, a dialogue scene has multiple starting conditions, and different starting conditions will lead to different directions in the dialogue process between the robot and the user. So in the first step of scene design, we need to sort out all the different starting conditions in the dialogue scene.

Example 3 Scene Check Weather

Conditional example

Example 4 Scene Leaving

Conditional example

2. Dialogue process

After sorting out all the initial conditions, you can write the process under each condition according to. For more information, see the following example.

Example 5? Scene departure

? Starting condition: only vacation is included.

? U: ask for leave tomorrow

? A: When?

? U: the day after tomorrow

? Please tell me the job number.

? U:00 1

? Your application for leave has been submitted.

Example 6

Dialogue start condition: only the number of vacation days is included.

U: please take three days off.

When did you start asking for leave?

U: the day after tomorrow

A: When will it arrive?

U: Friday.

Please tell me your work number.

U:00 1

Your application for leave has been submitted.

It can be seen from Examples 5 and 6 that different startup conditions will lead to different dialogue processes. So it is very important to sort out all the initial conditions. However, after combing the dialogue process, the part of the scene design is higher than the paragraph. Next, we need to clarify our intentions.

Three. Design of Dialogue Intention

This part mainly introduces the classification of intention, how to disassemble dialogue intention in the scene, and the design and optimization of dialogue intention knowledge.

Definition and classification of intention

Intention refers to the user's purpose. If the user says "I want to listen to music", then listening to music is the ultimate goal of the user. We just need to play him a piece of music. In the use of intelligent customer service, intentions are generally divided into two categories: question and answer intentions and dialogue intentions. Question and answer intention refers to asking and answering. For example, each merchant has a fixed answer to the questions of "time limit for refund" and "time limit for refund". This is the purpose of asking and answering.

However, in the dialogue scene, the user will interact with the robot many times, and each interaction, the user will express his purpose, which is actually what we call dialogue intention. Look at the example.

Example 7 Scene Check Fast

When we buy goods online, we will check the delivery of the goods we buy. The user in the picture says, "Where's my courier?" . The sentence "Where is my express?" It is a dialogue intention, and the user said that its purpose is to inquire about the progress of express delivery.

2. Dismantle the intention of the dialogue

A dialogue scene generally contains multiple dialogue intentions. Basically, every interaction between the user and the robot is a separate dialogue intention. We can write down the dialogue flow of Example 7 and have a look.

Example 8? On-site inspection express

Starting condition: multiple commodity orders

U: Where is my express?

Order list

U: select an order

A: it shows the order delivery.

In this scenario, the user expresses that one purpose is to "query the order status", then the robot will push the order list to the user to determine which order the user is asking about. This is the intention of a dialogue. Then, after the user confirms the specific order, the robot needs to send the logistics information of the confirmed order to the user, which is also a dialogue intention.

So the above scene contains two dialogue intentions. By this time, the dismantling of the dialogue intention of this scene is over.

3. Knowledge design of dialogue intention

After disassembling the dialogue intention contained in the scene, we need to design the intention. Let's use the scene leave as an example.

For example? Scene departure

? Starting conditions: only the starting time is known.

? U: ask for leave tomorrow

? A: When?

? U: the day after tomorrow

? Please tell me the job number.

? U:00 1

? Your application for leave has been submitted.

There are three rounds of dialogue, so there are three dialogue intentions in this process. Let's first look at the dialogue content of the first dialogue intention.

U: ask for leave tomorrow

A: When?

Ask for leave tomorrow, in this sentence, tomorrow is an essential word, representing the time when the leave begins. So this intention = the time when the leave begins, and its composition is a specific time/date, such as today, tomorrow, 65438+February 30, etc. The general dialogue system will have all the time words built in, so we won't do bad work here.

This dialogue only contains one element word, but the number of element words included in the dialogue intention varies according to different starting conditions, which can be set with reference to the flow of each starting condition.

4. Knowledge optimization of dialogue intention

Everyone expresses the same sentence differently. In order to make the robot understand the meaning expressed by users more accurately, it is necessary to optimize the knowledge. The optimization steps include: writing complete sentences-disassembling and combining-inducing synonyms.

Example: Where is the express delivery?

Where is the logistics?

Where are the things?

These words all mean "check logistics", but the words are not exactly the same. Through disassembly, we can draw the following conclusions: this knowledge is synonymous with express delivery+where to go.

Example:? Check express delivery

? Check express delivery

? Quick check

? Check the commodity logistics.

This group of sentences also means "check logistics", and their words are not exactly the same. Through disassembly, we can draw the following conclusions: this knowledge = synonym for check+synonym for express delivery.

By analyzing the above two examples, we can clearly see that the combination of these two groups has only two phrases, one of which is the same, so we can merge the two groups: express synonym +cha/where synonym?

Four. On-line management of intentional knowledge

1. robot assignment

After designing and optimizing all the question-and-answer intentions and dialogue intentions, we need to import these prepared knowledge into the online platform of the robot for operation. General customers only need to build a robot on the platform and bind multiple knowledge bases at the same time, which can achieve the effect.

But like some e-commerce customers with large passenger flow, they will build robots according to the whole business scene, such as pre-sales robots, sales robots and after-sales robots. One advantage of this is that when the customer asks some questions that the robot can't answer, he can directly transfer to the relevant category of seats.

There are some customers, which will be more complicated. They are a big enterprise, with many departments inside, and each department is using the same background.

For example? A company has human resources, finance, administration, production, projects and other departments, and they use the same enterprise account to log in to a backstage. Every department will establish a knowledge base, but if they are all bound to a robot, some problems will arise.

If the same enterprise account must be used, in this case, it is suggested that each department can set up a robot in the background, and only bind the robot of this department to the relevant knowledge base of this department. (It is best to have permission settings) However, it is best for each department to use an independent enterprise ID and log in to their own background, which will not cause interference.

2. Evaluate and adjust the robot.

After a period of operation, we need to know how effective the robot is. Generally speaking, as long as the dialogue log of the robot is exported and analyzed.

1) The content of log analysis.

The general analysis log will analyze the following contents

Total number of interactions, total number of completed interactions, number and proportion of unidentified intentions, number and proportion of identified false intentions, and invalid intentions.

The total amount and proportion of question and answer intentions, the total amount and proportion of dialogue intentions, and the proportion of dialogue intentions and question and answer intentions to the total amount.

The intention and proportion of high-frequency question and answer, the intention and proportion of high-frequency dialogue, and the quantity and proportion of dialogue scenes.

2) Adjust the robot

After the analysis, the intention and knowledge of the robot will be adjusted and optimized according to the specific results.

Problems needing attention

There are several possibilities for unidentified situations with many problems.

(1) The user said that there is no similar problem in the knowledge base, so please add it.

(2) The user said that there is a similar problem in the knowledge base, and you need to manually check whether it is an error caused by other knowledge or technical reasons. ......