Chatbot Training - Detailed How-To Guide For 2022

Post by
Indrek Vainu
May 5, 2021
Chatbot Training - Detailed How-To Guide For 2022


For proper chatbot training you should start by making sure your chatbot topics do not have significant overlap. They need to be different enough that the bot can decide where to pass incoming customer queries. Example of well built topics - “credit card” and “debit card”. Example of poorly built topics - “credit card gold for professionals” and “our new gold credit card”.

Keep your chatbot training data clean. Make sure to have at least 10 phrases trained for each chatbot topic (aka intent). Make sure the phrases trained for intents actually reflect that intent. Do not include phrases that could be associated also with other phrases. 

1. How To Train The Chatbot

Let’s say you have a topic “My webinar”. Now we want the bot to be able to answer with this topic when users will come and ask about the next webinar. 

To do that, we need to train the topic. Click on the “Train” tab above the chat window. To build along, you can create a free account here

Chatbot training with

This opens up the training section for the topic. Here all you need to do is enter a few training phrases. Phrases are different ways of how people might ask about the “My webinar” topic.

We recommend adding at least 5 different phrases. With less than 5 phrases, the AI will have a bit too little training data. There is no upper limit on the number of training phrases you can add. 

However, keep in mind that the phrases need to be relevant to the topics. For example, no need to train the phrase “How much does your service cost” to the topic “My webinar”. For that phrase, just make a new topic called “Service cost” and train the phrase there.

Chatbot training with

And now that you have your training phrases inserted click the “Train” button in the left side menu.

Conversational AI training with

This opens up the Train menu. Here you see your “My webinar” topic with the phrases that you entered. All you have to do now is click “Train” on the top right and let the AI train itself. After a minute or so it will complete its job.

AI training with

2. Topic Rating Matrix And What To Keep In Mind When Building Topics

It is important to decide which topics are important for the user and the business goals. But not less important is also to understand the overall solvability of every topic. For this reason, we have developed an evaluation tool called Topic Rating Matrix (TRM). It is a simple but very useful tool that helps to figure out what should be added to the intelligent virtual assistant’s repertoire and what should be avoided.

Topic Rating Matrix for chatbot training

List out 30-50 most frequently asked topics that your customers ask via messaging. Then for each topic give a score of 1-3 for occurrence and 1-3 for solvability. For occurrence score as follows - give 1 where occurrence is low, 2 when it is medium and 3 when the topic gets asked a lot.

For solvability give the topic a score of 1 when only a human customer service agent can solve the customer questions. Give 2 when the bot can solve it with additional information (e.g. authenticated backend query, information on past customer behavior). Give a score of 3 to the topic when the bot can solve the issue on its own from a common FAQ dataset.

Next, multiply the occurrence and solvability scores. Now you get a list which ranks the topics by impact taking into consideration the volume of inbound queries and the actual ability for the AI to solve the issue.

Based on this list, pick the top 20 or 30 and build those into your chatbot first. These are the low hanging fruits that you should train first with the bot. 

In a telecom intelligent chatbot, for example, many customers might have problems with mobile connection, but there can be a number of reasons for poor connection. So the occurrence is high but the solvability is tricky. In a nutshell, one should aim to include topics that have high occurrence and high to medium solvability. 

An optional step is to organize your potential topics first in a spreadsheet like this. Feel free to talk also with your team members to find the topics that should be prioritised the most to get additional input into the selection process of which topics to train first.

3. Conclusion

There you have it! You now have a framework of how to select topics for training through the Topic Rating Matrix. In addition you see a detailed process on how to train the chatbot and how to think about chatbot training phrases.

If you are looking for AI chatbots for your customer service, then feel free to sign-up for an account with In just 5 minutes you can get your own natural language understanding AI chatbot that you can connect with your website.