Note: This feature is currently available as part of an Early Access Program (EAP). Please reach out to your CSM to enable it on your account.


Understanding the kind of utterances and intents we could expect from our customers and training our bots to respond to these accurately is essential. Following are a few best practices to follow when setting up intents:

  1. Maintain different variants of utterances 

Changing only letter cases, grammar, numbers, names, and places will not give the model enough variation in training data. It will lead to bots looking for the exact words in the customer chat, leading to false positives. Intents can accommodate variations in grammar and spelling.

 

Incorrect

‘How to cancel the order?’

‘How to Cancel the Order?’

‘HOW TO CANCEL THE ORDER?’

 

Correct

‘How to cancel the order?’

‘I don’t want this product.’

‘Where can I cancel my order.’


Incorrect

‘How to book a flight for the 25th of august?’

‘How to book a flight for the 26th of august?’

‘How to book a flight for the 27th of august?’


Correct

‘How to book a flight?’

‘Help me get a flight.’

‘Reserve a flight ticket.’

 

  1. Do not create multiple intents with similar customer goals. 

Instead, group these under the same intent. Similar intents can trigger an incorrect flow for the customer since it may not be apparent to the bot which intent is relevant.


Incorrect

Intent: Purchase health insurance

Intent: Purchase life insurance

Intent: Purchase travel insurance


Correct

Intent: Purchase insurance


This intent can then be mapped to a flow that requests the kind of insurance to purchase


  1. Do not have the same utterances as part of different intents. 

This can confuse the bot on the flow to present and may decrease the reliability of the response.


  1. Retrain utterances and questions that are unmapped or unhelpful. 

The admin must attempt to constantly link the unmapped customer questions to an answer, add variants to improve coverage, and check unhelpful responses to rectify them. The more unmapped customer questions are trained by the admin, the more the machine learning model’s performance improves.


  1. Add complete sentences to provide the full context to the bot for training.


Incorrect

‘Cancel order’

‘Stop shipping’


Correct

‘How to cancel the order from shipping?’

‘How to stop the shipping of the purchased order?’

  1. Add as many utterances to intents to train them better

While the minimum required number of utterances is 5, more variations will lead to more accurate bot performance.

 

  1. Try adding a similar number of utterances to each of your intents.

Say you have intents titled Purchase, Return, and Cancellation, and each has 5, 15, and 50 utterances, respectively.  An unequal number of utterances can lead to better accuracy for specific intents over others. It is best to keep the number of utterances similar across intents.

 

  1. It is not necessary to maintain utterances in multiple languages

The bot can translate utterances even if you maintain them only in the primary language configured in your account. Maintain multilingual utterances if you want an exact match or if specific text patterns need to be established.


Other helpful articles


Configuring intent detection for bots

Retrain Natural Language to improve bot performance