A guide to chatbot structure

I wrote an article "​​How to Design a Chatbot More Elegantly​" a few days ago. Some friends left a message and asked me: Stone, are there any articles about chatbot architecture instructions? ? Where there is demand, there is motivation. Today we will talk about the architecture of chatbots.

Today, more and more enterprise customer service systems (and of course other business systems) are shifting from traditional voice calls to text, graphics and intelligent voice.

#Communicating via chatbots is becoming more and more popular for two main reasons: simplicity and real-time.

Below, let’s talk about how chatbots work, how to customize them and everything you need to know about chatbot architecture.

But before we get started, let’s cover the basics.

What is a chatbot?

A chatbot is a program that simulates conversations between people and computers, or between people. When asked a question, the chatbot responds using a knowledge database.

Artificial Intelligence (AI) is used to simulate natural language conversation or chat. Common ways are via messaging platforms, mobile apps or phone calls.

Chatbots enable communication between humans and machines, work independently of human assistance, and use technologies such as natural language processing (NLP) to answer questions. Natural language processing (NLP) is a branch of artificial intelligence that enables computers to understand text and spoken language in much the same way humans do.

How does a chatbot work?

Chatbots allow users to easily find answers to questions and question requests via text, audio, images, and more without human intervention.

Chatbot is an automated solution that allows businesses to handle multiple customer inquiries simultaneously. According to some statistics, most customer service services absolutely need to be available 24*7 hours a day.

Now most enterprise chatbots have integrated more rules and natural language technology, and the latest models can continuously learn during use.

Today’s AI chatbots use advanced artificial intelligence tools to clarify the true purpose of the customer.

There are two main types of chatbots, as shown below.

Rules-Based Chatbots

These types of bots can only understand a limited number of options that they have been programmed with. Has the following advantages:

• Easy to build: Use true and false algorithms to understand customer queries and come up with relevant answers.
• Easy to implement: There is no need for high learning costs, and it may be possible to implement it with only simple keywords or regular expressions.
• Easy to control: The rules are set by the enterprise itself, so the output answers will not exceed the set range.

Of course there are advantages, but there are definitely disadvantages:

• Strong dependence: too much reliance on rules, beyond the predefined rules, and unable to understand their meaning
• Menu-based operations: During the interaction, the chatbot displays a series of options that the user needs to choose from, which makes it very difficult to truly understand the user’s true intention as it may not be represented in the options.

Chatbots based on artificial intelligence

These chatbots are relatively complex, adding artificial intelligence algorithms to the original ones. Use natural language processing (NLP) and semantics to respond to open queries. AI chatbots can recognize language, context, and intent and respond accordingly. is a more complex chatbot.

In this space we have found two different approaches:

Probabilistic Chatbots

This type of bot uses end-to-end machine learning to create history-based A model for conversation logs, rather than through intent detection or finding relevant responses in a knowledge base. Although they do not follow a fixed script and can be interacted with naturally, probabilities also have disadvantages:

• When they learn from experience and data in conversations, a lot of bias can be introduced. There is limited control over the output dialogue, and it is possible that the bot will give some controversial answers and receive complaints from customers.
• Implementing a probabilistic chatbot requires a large amount of training data. The more data obtained, the better its accuracy. This is a painful and long work for the developers who collect the data.
• The answer the chatbot makes is in a "black box" (model), which means how the chatbot makes the answer. There is no transparency, and it is difficult to modify or adjust the inference results.

Deterministic Chatbot

This kind of chatbot uses natural language processing to calculate the weight of each word, analyze the context and meaning behind them to output a result or answer.

These chatbots are able to match intent to answers based on meaning.

They have their advantages and disadvantages:

• Only output the content filled by the company, it is easier to control the reply tone and corporate image.
• This is not based on probability learning, which can prompt new hot topics to include.
• Follow a deterministic decision tree to guide the client to the desired outcome. Decision trees can be very complex and are overseen and controlled by a trainer who will not accept controversial, unpopular answers.
• Whenever there is no relevant content in the knowledge base to respond to the user, the trainer can retrain the model or formulate rules, thereby achieving a smooth transition and reducing basecase.

Friends who are considering introducing a chatbot can learn about the chatbot architecture, which can combine all content together. Of course, you also need to master automated testing.

What is chatbot architecture?

The architecture of a chatbot depends on its purpose

No matter which chatbot you use, the robot communication process is basically the same.

Programming languages ​​Java, Python, PHP and other languages ​​can be used to create bots that respond to queries. Most conversations start with a greeting or question and then lead the user through a series of questions. to get the answer.

The following is a detailed introduction to the basic architecture of the chatbot.

Natural Language Understanding Engine

This is the core and most important first step. The user enters a message and NLU reads the message to understand the user's intent. The rules engine then starts calculating the best response.

You need to spend some time thinking about your QA collection library, and collect the QA library logically and regularly. Of course, you also need to understand the QA testing strategy.

Knowledge Base

This is a base of information about products, services or business needs. It can include FAQs, troubleshooting guides, information about services, or how to do business.

Both knowledge and databases provide the chatbot with the information it needs to respond authoritatively to the user.

Data Storage

This is where analytics and conversation logs are stored. As chatbots are used longer, more specific and complete analysis solutions need to be developed to make the models more accurate and cover wider.

At each stage, the business must be systematized to ensure that the chatbot is connected with the business.

What architecture is needed for the most basic chatbot?

Small businesses and marketing campaigns often start with a level one chatbot. These can usually only be built on one platform. This category excels at handling simple problems that make up 70-80% of common problems. This type of chatbot answers simple questions, such as "What time will you open?"

When users require more complex information (such as problem diagnosis), the chatbot needs to be scaled up.

For example, if someone asks: "What's wrong with my delivery?"

This will require a higher-level chatbot.

As the capabilities of chatbots become more intelligent and the business they can handle becomes more complex, more traffic exposure is required

HTTP and chat interface

2 Level chatbot is semi-scripted and features a live chat widget. Here you can chat with the customer support team directly from the home page.

Message Broker

This is where the publisher (such as a chat interface) adds messages to the queue. Customers access chatbots through instant messaging platforms such as WeChat, DingTalk, Enterprise WeChat and QQ.

Live Broadcast Agent Platform

If the robot fails to correctly identify the user’s intention, the human agent can seamlessly intervene. In some cases, they will resolve the issue and hand the end of the conversation back to the bot.

The bot can also call up customer details from the Customer Relationship Management (CRM), such as changing a password or looking up an order.

Enterprise-Grade Architecture

Taking chatbots to the next level requires the use of technology to enable complex conversations. You also need to determine how to extend the functionality of your software.

Of course, every business is different. Here’s a summary of some common technologies, workflows, and patterns needed to build bots with enterprise-grade architecture.

There are many considerations beyond core functionality. A software test scheduler must be built into any chatbot of choice.

A conversational robot can be divided into a "brain" and a set of requirements or "modules".

How Chatbots Work

Chatbots work using three classification methods:

• Pattern Matching
• Algorithms
• Neural Network

Pattern Matcher

Bots use pattern matching to analyze text and generate appropriate responses. The standard structure of these patterns is Artificial Intelligence Markup Language (AIML), you can refer to iFlytek's "abnf Grammar Specification"

For example:

Qiao ·Who is Biden? .

The chatbot knows the answer because his or her name is part of the relevant pattern. But for more advanced information beyond relevant patterns, chatbots can use algorithms.

Algorithm

The algorithm reduces the number of classifiers and creates a more manageable structure. In the following example, each term is assigned a score.

Input: "Hello, good morning."

Term: "Hello" (no match)

Term: "Good" (Category: Greetings)

Term: "Morning" (Category: Greetings)

Category: Greetings (Score = 2)

With the help of scores, one can find word matches for a given sentence, This identifies the category with the highest matching degree.

Natural Language Processing Engine

This engine uses weighted connections to calculate input and output. Each step used in the training data modifies the weights to improve accuracy. Sentences are broken down into individual words, and each word is then used as input to match the content of a network database. Then keep testing the words.

Additional considerations for enterprise-level architecture

In addition, chatbot architecture must also consider the following elements.

Security

Security, governance and data protection are to be taken seriously. This is especially important for businesses that store information about millions of customers.

If users do not want their personal details to be leaked, they need to consider how to remain anonymous. If you want to access personal information, you need to do so in a secure manner.

It is important to establish confidentiality measures so that no one can gain unauthorized access to sensitive systems.

Any small mistake, such as a spelling mistake or a broken hyperlink, has the potential to be seen by thousands of users every month.

A small mistake can have a huge impact on your business image.

Summary

Chatbots simplify interactions between people and services, thereby enhancing customer experience. They also provide businesses with the opportunity to improve re-engagement processes while reducing customer service costs.

The above is the detailed content of A guide to chatbot structure. For more information, please follow other related articles on the PHP Chinese website!