Building the NLU model

A workshop should be held around the purpose. The business often knows where the issue of the internal processes. It could be that the customers have recurring questions. Once you have formed an idea of what the chatbot could help with it is time to start building. In our example, we will focus on a dialogue-based chatbot that helps answer the recurring repetitive questions of the business.

When the chatbot receives questions about a specific topic or question, we can choose to create an "Intent" and a corresponding "Workflow" for the question or topic.

An "Intent" can be explained as a trigger that contains training data in the form of potential sentences or questions that can be asked by a customer. The intent is what the bot will try to match for when a customer asks a question and if the bot gets a match against an intent, the answer we linked to the intent will be played back. This answer is called a “Workflow”.

A "Workflow" can be explained as a small process where one or more steps are played back in an order that is from top to bottom. The flow can include text responses, counter-questions, calculations, time activations, and more.

Creating an intent

The button creates a new intent. When you click this button a popup appears and asks for a name. Give an appropriate name for the intent so that it is easy to understand. Eg: Category-subject or Bus-pass application.

Once we have created our intent we add training data based on what the customers may respond to. This can be edited as you gain more experience with your bot.

It is usually enough with 5-10 different examples.

When you’ve written a training sentence, press enter to store it in the training model.

Sometimes, several intents are created during the training session. Just repeat the steps above until you are done creating all your intents.

When you're done creating your intents and you've filled it with training data, you need to train your NLU(nature language understanding) model. To do this, click on the "train model"  button.

The training engine will train the model in a predetermined number of cycles to optimise the NLU model. A small NLU model may take a few minutes while a larger model can take a longer time. Therefore it is recommended to train after building or after training but not during building.

If you haven't finished training data for the intent you're working on, you can choose to exclude it from training cycles. You can do this by clicking on "Active". When you're done with your intent and activate it again as before.

You can change the number of training cycles in the settings menu. When you reduce the number of cycles you also lower the quality of how well your bot will respond. However, your bot will complete the training faster. This can be good when you are starting out building your bot. Once you have completed a major change in your bot, you should do a training with at least 200 to 300 training cycles to increase the quality “control” (if this makes sense) of the NLU model.

Once you have completed your intents building session, you need to create workflows, i.e. answers to be published when an intent is stuck.
You do this under ”workflows” in the main menu.

1. In the main menu, go to “Intents”
2. Click the blue “plus” button to add a new intent.
   1. Name your intent “Demo-Exercise”
3. The system automatically selects your new intent and you are ready to give your intent some training data
4. In the textbox “Add Training Sentence”, add 5-8 question sentences similar to the focus of your intent
   1. “Is this a demo?”,
   2.  “Can you show me a demo?” and so on. Feel free to improvise but try to get the feel of how your customers would ask the question about the subject. After each new added sentence you will see a notification saying “Model saved successfully”
5. When you are finished with adding your training data you are ready to run the training cycles to implement your new intent in the NLU-model.
6. When you have created all of your intents, click “Train model”  to train your NLU-model.