top of page

Chatbots and how they work

Updated: Feb 13, 2020

From the customer support dialog boxes you find on e-commerce websites to virtual assistants like Siri and Alexa, it’s likely that you’ve encountered chatbots frequently in your everyday life. Like its name suggests, a chatbot is a piece of software designed to conduct a conversation or dialog. They can be found in a wide range of industries to serve a variety of purposes, ranging from providing customer support to aiding in therapy to simply being a source of fun and entertainment. Let’s take a closer look at how chatbots are able to do what they do!

How chatbots work

To interact with the human user, chatbots must be able to:

parse the user inputinterpret what it meansprovide an appropriate response or output

For example, a user query could be, “Show me hotels in Los Angeles for tomorrow.”

A good chatbot will be able to identify the intent and entities of the query. The intent is the purpose or category of the user query, such as to retrieve a list of hotels. Entities are extra information that describes the user’s intent. In this case, the entities are “Los Angeles” and “tomorrow.” With these pieces of information, chatbots should be able to respond to the user with a list of available hotels for the correct location and date.

Types of chatbots

Chatbots generally fall into a few broad categories, depending on what purpose they are designed to serve.

Response architecture models

First, chatbots can be categorized according to how they generate the response that gets returned to the user. The simplest approach is the rule-based model, where chatbot responses are entirely predefined and returned to the user according to a series of rules. This includes decision trees that have a clear set of possible outputs defined for each step in the dialog.

Next, there is the retrieval-based model, where chatbot responses are pulled from an existing corpus of dialogs. Machine learning models, such as statistical NLP models and sometimes supervised neural networks, are used to interpret the user input and determine the most fitting response to retrieve. Like rule-based models, retrieval-based models rely on predefined responses, but they have the additional ability to self-learn and improve their selection of response over time.

Finally, generative chatbots are capable of formulating their own original responses based on user input, rather than relying on existing text. This involves the use of deep learning, such as LSTM-based seq2seq models, to train the chatbots to be able to make decisions about what is an appropriate response to return.

While generative models are very flexible and powerful in that they are not confined to a predefined set of rules or responses, they are also significantly more challenging to implement. Training these chatbots require an abundance of data, and it is often unclear what gets used for their decision-making, making them more prone to grammatical errors and nonsensical replies. By contrast, retrieval-based models can guarantee the quality of the responses since they are predefined, but these chatbots are in turn restricted to language that exists within the training data.

Thus, chatbots often use a combination of the different models in order to produce optimal results. For example, a customer support chatbot may use generative models for creating open-ended small talk with the user, but then are able to retrieve professional, predefined responses for answering the user’s inquiries regarding the business or product.

Conversation domains

Chatbots can also be categorized based on the range of conversation topics they are able to cover. Closed domain chatbots, or dialog agents, are restricted to providing responses with a particular focus, such as booking a hotel room. Because they are designed with a specific goal in mind, these chatbots are often very efficient and have great success in accomplishing what they are intended to accomplish. The user-perceived quality is also high, because users don’t expect the chatbots to provide responses outside of the pre-established domain.

On the other hand, open domain chatbots, or conversational agents, are capable of exploring any range of conversation topics, much like how a human-to-human interaction would be. Many of these “companion bots” have filled the roles of a friend or therapist, allowing the user to connect with them on an emotional level. While they have great potential, open domain chatbots are challenging to implement and evaluate.

It is worthwhile to note that the term “chatbot” is sometimes reserved only for open domain conversational agents, but for the purpose of this course, we include closed domain dialog agents as well.


Another way chatbots can be categorized is by which side – the user or bot – is able to take initiative on the conversation. Looking back at our previous example on hotel search, notice how the user is free to provide their request in their own words, and the chatbot is able to identify and piece together the relevant keywords to answer the query. This is an example of a mixed-initiative system, representative of a normal human-to-human conversation where all participants have the chance to take initiative.

By contrast, a system-initiative system is one where the chatbot controls the conversation and explicitly asks for each piece of information, such as the date and location for the hotel booking. While this system is more straight-forward to implement because user response can be anticipated, it lacks the flexibility and naturalness that characterize a normal human dialog.

Into the future

Chatbots have come a long way. What started out as computers that attempted to mimic human conversation has grown into elaborate systems that are able to carry out a multitude of functions and goals.


It has 3 specific concepts like:

Entities: Entity basically represents a concept in your Chatbot. It might be a payment system in your Ecommerce Chatbot.

Intents: It is basically the action chatbot should perform when the user say something. For instance, intent can trigger same thing if user types “I want to order a red pair of shoes”, “Do you have red shoes? I want to order them” or “Show me some red pair of shoes”, all of these user’s text show trigger single command giving users options for Red pair of shoes.

Context: When a NLU algorithm analyzes a sentence, it does not have the history of the user conversation. It means that if it receives the answer to a question it has just asked, it will not remember the question. For differentiating the phases during the chat conversation, it’s state should be stored. It can either be flags like “Ordering Pizza” or parameters like “Restaurant: ‘Dominos’”. With context, you can easily relate intents with no need to know what was the previous question.


For each kind of question, a unique pattern must be available in the database to provide a suitable response. With lots of combination on patterns, it creates a hierarchical structure. We use algorithms to reduce the classifiers and generate the more manageable structure. Computer scientists call it a “Reductionist” approach- in order to give a simplified solution, it reduces the problem.

Multinational Naive Bayes is the classic algorithm for text classification and NLP. For an instance, let’s assume a set of sentences are given which are belonging to a particular class. With new input sentence, each word is counted for its occurrence and is accounted for its commonality and each class is assigned a score. The highest scored class is the most likely to be associated with the input sentence.

For example Sample Training set

class: greeting “How you doing?” “good morning” “hi there”

Few sample Input sentence classification:

input: “Hello good morning” term: “hello” (no matches) Term: “good” (class: greeting) term: “morning” (class: greeting) classification: greeting (score=2)

With the help of equation, word matches are found for given some sample sentences for each class. Classification score identifies the class with the highest term matches but it also has some limitations. The score signifies which intent is most likely to the sentence but does not guarantee it is the perfect match. Highest score only provides the relativity base.

Artificial Neural Networks

Neural Networks are a way of calculating the output from the input using weighted connections which are calculated from repeated iterations while training the data. Each step through the training data amends the weights resulting in the output with accuracy.

As discussed earlier here also, each sentence is broken down into different words and each word then is used as input for the neural networks. The weighted connections are then calculated by different iterations through the training data thousands of times. Each time improving the weights to making it accurate. The trained data of neural network is a comparable algorithm more and less code. When there is a comparably small sample, where the training sentences have 200 different words and 20 classes, then that would be a matrix of 200×20. But this matrix size increases by n times more gradually and can cause a huge number of errors. In this kind of situations, processing speed should be considerably high.

There are multiple variations in neural networks, algorithms as well as patterns matching code. Complexity may also increase in some of the variations. But the fundamental remains the same, and the important work is that of classification.

Recent Posts

See All

Generative AI report

Top GenAI companies: OpenAI Google Anthropic Meta Mistral Stability AI MidJourney Top GenAI Models GPT 4 Gemini 1.5 Llama2 Mistral Claude Stable Diffusion


bottom of page