Genesys Cloud - Main

I am doing something similar and a question i had was how does it decide which to handle first. is it the order the intents are identified? in your example Change of address would be first. or is it the order that the intents were created? or is it the intent with the most health/utterances?

@Anton, i had one intent with multiple different types of slots and used a switch, i had an issue with no matches as sometimes it would go down the default path of the switch and that wasnt helpful. It was also difficult to come up wording when a confirmation prompt was presented. also, if one slot was chosen for the intent and the customer came back, and did another utterance with mutilple slots, it would find the slot that was already selected, it wouldnt update the slot - i did try the Clear slot function, but it didnt seem to work. at that point i decided i needed to break my 1 master intent into 3, rather than trying to use 1 intent with 3 different slots.

also wondering how others handle multiple intents

------------------------------
Kellyn DuBois
Lumen Technologies Service Group, LLC
------------------------------

Its not something easily handled, and generally I'd say to stay away from it. A few things to consider:

Do you have true proof that this is a problem that needs to be solved? Ie are people actually saying multiple intents often enough its worth investing in?

The second thing, is that someone could also equally likely (though your specific data would prove this true or false) something like "I just changed my address, now I want to pay my bill" and unless you have the greatest intent training of all time, you would likely end up with the intent algorithm assuming they want to change their address, even though they said they just changed it. As you may or may not know, the intents only know the world theyre trained on, so unless you're also training on "already changed address" as an intent there'd be obvious confusion. 

That's a fairly specific examples, but theres many other ways that trying to understand more than one intent can be equally hurtful as helpful when its real. 

IMO the ideal way to handle is this is simply through disambiguating between top intents. In your example you'd have relatively high confidence for both 'change address" intent AND "pay my bill" intent. Let the user choose which to do first, or build a single intent to cover both if thats a common combo, which would force them to change their address before paying the bill, if thats your business process. Then, once they finish the first one, they will naturally follow up with the second thing they want to do. If, as I mentioned before, these two are a specifically important combination that is frequent and real, then use a single task to handle it end to end (similar to the comment below) 

If you're worried about any and all combinations of your intents, its likely too complex of a web you'd weave to make it worth it. 

As a semi-related note, Ive built many bots with many clients, and what I see happen time and again is that they build hundreds of intents, and in many cases this can actually degrade accuracy, and over time the data gets quite messy. We would eventually have a session where we review their bot, and align it with their data, and realize that something like 30% of their intents are never hit, or less than 5 times per month. Which really means you have a lot of bloat in your build and it just adds more opportunity for confusion. That's why I started with my first sentence, first make sure this is a real problem occuring that will truly increase your KPIs. And if it is, then the solutions above hopefully help you start. 

I have come to some of the same conclusions as I have built my first few bots now. If there isn't a way to capture the multiple intents, is there any way to weight utterances/intents differently? If I have a bot asking for medical symptoms in order to immediately direct a caller to the appropriate level of care, it may have utterances such as "cut my finger" and "left arm is numb" on a couple of different intents. If I can't capture that the caller said both of those, is there a way that I can treat one of them as more important that the other? I don't care if the NLU score for "cut my finger" is 0.95 and the score for "left arm is numb" is 0.75. I still want "left arm is numb" to win.

Its not something easily handled, and generally I'd say to stay away from it. A few things to consider:

Do you have true proof that this is a problem that needs to be solved? Ie are people actually saying multiple intents often enough its worth investing in?

The second thing, is that someone could also equally likely (though your specific data would prove this true or false) something like "I just changed my address, now I want to pay my bill" and unless you have the greatest intent training of all time, you would likely end up with the intent algorithm assuming they want to change their address, even though they said they just changed it. As you may or may not know, the intents only know the world theyre trained on, so unless you're also training on "already changed address" as an intent there'd be obvious confusion. 

That's a fairly specific examples, but theres many other ways that trying to understand more than one intent can be equally hurtful as helpful when its real. 

IMO the ideal way to handle is this is simply through disambiguating between top intents. In your example you'd have relatively high confidence for both 'change address" intent AND "pay my bill" intent. Let the user choose which to do first, or build a single intent to cover both if thats a common combo, which would force them to change their address before paying the bill, if thats your business process. Then, once they finish the first one, they will naturally follow up with the second thing they want to do. If, as I mentioned before, these two are a specifically important combination that is frequent and real, then use a single task to handle it end to end (similar to the comment below) 

If you're worried about any and all combinations of your intents, its likely too complex of a web you'd weave to make it worth it. 

As a semi-related note, Ive built many bots with many clients, and what I see happen time and again is that they build hundreds of intents, and in many cases this can actually degrade accuracy, and over time the data gets quite messy. We would eventually have a session where we review their bot, and align it with their data, and realize that something like 30% of their intents are never hit, or less than 5 times per month. Which really means you have a lot of bloat in your build and it just adds more opportunity for confusion. That's why I started with my first sentence, first make sure this is a real problem occuring that will truly increase your KPIs. And if it is, then the solutions above hopefully help you start.