We have our Chatbot in testing and I’m adding intents but I wondered if anybody can advise on best practice or proven approach.
I recall from the Chatbot training that we were advised to have about 10 utterances for each intent but I seem to have anything from 3 to 30 utterances for each intent. I don’t want to add too many and it slow things down but I don’t want customers to not be able to find what they are looking for.
So, if anybody has experience they’d be prepared to share that would be very much appreciated.
There is a line between over- and undertraining a natural language model. Too few, and it will be very precise in what it is looking for, and too many, and it could overtake all the intents.
For this reason, we recommend around 10 utterances to help the system understand the intent.
The key thing to note is that you do not have to add exactly what the person is going to ask for the intent to be recognised. It will use its natural language understanding to process and best match the intent required.
I personally use AI (like Chat GPT, Co-Pilot, or Gemini) to help me come up with some utterances that people may ask for my query.
For example, I have an Intention to Book a holiday. I have added 10 utterances that people may ask for. I asked an AI to help me think of 10 ways a person could ask this question and I used that to help me fill out the spaces.
Then, when I go to test, I can use a sample group of people and ask them to type what they require. I can then see what the intents match and if I need to add their utterances to my Intent.
This is always an ongoing process of checking what utterances people are asking for and then adding them to an Intent, ignoring them or adding to an Unknown intent option.
