AI model is the brain of your bot. The better an AI model is, the better your bot can understand customer queries.
A new AI model is created automatically every day at midnight if any changes have been made to the training data that day. However, an initial AI model has to be set up in order to kick off this process.
Follow the steps below to create the initial AI model for your bot:
Step 1: Identify most frequent intents
A good intent structure is the foundation of a strong AI model. This means intents are based on real customer data, clearly defined and the expressions trained to them do not overlap with other intents.
Depending on your available data, language and industry, use Content Coverage Analysis or the Impact Report to identify the top intents and start adding expressions. You can use the Impact Report or Training Center to identify expressions.
At this stage, you may only have 10 meaningful intents with a handful of expressions - try an aim for 15 good expressions per intent for now. You will need more to launch the bot, see What is the optimal amount of expressions to learn more.
Step 2: Create the initial AI model
Now you have a basic intent structure, you want to create the initial AI model.
Go to AI > Models > +New Model > Train Bot.
When generating a model, there are a number of settings you can select or unselect:
- Run Cross-validation: This will generate an evaluated model which is automatically created once a week, whether training has taken place or not. A Confusion Matrix will help you identify at this early stage how clearly defined intents are and if there is overlap between intents.
Advanced Model Settings
Advanced Model Settings affect the way the AI model works. There are four key settings to look at:
Out of scope
Out of scope is a way for the AI model to deal with unknown topics. By default, the AI model assumes that every customer message can be assigned to an existing intent. This is, of course, not the case. For example, if a customer asks an eCommerce bot ‘How do I order a pizza’, this question will not be covered by any intent.
-
- If selected, the out of scope acts as a hidden intent that the AI model can use when it cannot confidently predict the customer message to any other intent. The default reply is given when the out-of-scope intent is triggered.
- If unselected, the AI model will assume every customer message can be assigned to an intent
Unselected is the default, however is it recommended to set it to true.
Remove punctuation
- If selected, all punctuation will be removed.
- If unselected, data will remain it it’s original form
Unselected is the default setting but is it recommended to select this so the AI model does not consider punctuation when reading data.
Remove stop words
Stop words are words that are very common and may not carry much useful information for the AI model [e.g. ‘i’, ‘me’, ‘my’, ‘myself’, ‘mine’, ‘and’, ‘but’, ‘if’, ‘or’, ‘because’, ‘didn’, “didn’t”, ‘doesn’, “doesn’t”]
-
- If selected, all stop words will be removed
- If unselected, data will remain it it’s original form and stop words will be kept
Unselected is the default and recommended setting.
Lowercase input
-
- If selected, all data will be forced into lowercase.
- If unselected, data will remain its original case.
Selected is the default and recommended setting.
Once you click “Train Bot” on the very top of the Dashboard, the model can take several hours to generate. Check the status, which will move from Queued > Started > Finished.