NLU virtual agent topic

mehak1890
Tera Expert

Dear All,

I have some basic questions about NLU. I have enabled the NLU and i have created a Model/ intent then i have created the 6 utterances and also created same 6 test cases under test model and i have not created any entities under same intent.

 

mehak1890_0-1729938181633.png

 

Is it working fine or not? I'm not sure , but when I'm searching how to request the incident/ or problem on chatbot , the topic immediately starts, let's assume it's working fine. But when I'm searching another request in chatbot "I want to raise a request for phishing". The same topic "report a problem" is starting immediately, you know when I'm starting a new request then it should give me phishing topics or knowledge articles for phishing related if any in servcienow correct? which is not under the utterances what I understand. Can anyone pls help me how i can correct my utterances? so that if i'm searching " i want to raise a request for phishing" then some other results should come in chatbot?

 

Thanks,

Mehak

1 ACCEPTED SOLUTION

@mehak1890 

When it triggers unwanted intents "report a problem" for certain utterances "give feedback", then this is generally caused by one of the following - 

1. No vocabulary defined for acronyms or business terms used in the NLU Model, hence the NLU Engine can return unexpected predictions because it requires a synonym that expands the acronym and its meaning.
2. Samples utterances that have no context and consist of one or two words. Ensure all utterances have full context for the intent and avoid repeating common words that are not unique to the intent. AI does not like repetition, as it will skew predictions.

 

I see this issue time and time again. Hence, we introduced LLM topic discovery in Virtual Agent, which does not require the creation or maintenance of any NLU model, and we also provided tools for Migrating NLU/keyword Virtual Agent topics to LLM topics.

 

Hope this helps.

 

Regards,

Brian

View solution in original post

6 REPLIES 6

NLU Models are a painful experience, tweak it here and it goes wrong over there, tweak it there and something else doesn't work the way you want it, the LLM model is so, so much better and recommended if you have Now Assist. 

***If Correct/Helpful please take time mark as Correct/Helpful. It is much appreciated.***

Regards

Paul

Hi @mehak1890  - I'm so sorry for the delayed response and it seems like Brian's answer might be your solution.

In troubleshooting AI Search v Topics, it's not a bad thing for your phishing statements to go to fallback if your KB or catalog item content resolves the users question. This is why (I believe) we removed the Catalog Request topic - requests for KB content or items should be managed by AI Search and that frees you up to focus the topics on automation tasks that have very specific actions. "I want to request..." is a very generic phrase that could direct you across all your content. Given than the NLU VA is more linear, I find it helpful to balance my actionable automation tasks (topics) against content (AI Search). 
In your example, "I want to request a pizza", the actionable task is to enter size, toppings and delivery address -this could be a record producer (so it would go to AI Search - hopefully genius results) or it could be an actionable topic if you integrated your ordering system.  If it's an actionable topic, then all your example utterances are going to be variations of ordering, delivery, pizzas, pie and calling out vocabulary so the VA can recognize "Please deliver a large pizza with pepperoni to my house".  With the actionable topic, you're making the order process easier. 
That's not the same process if you were asking "I want to know the toppings for a pizza".  More importantly, you want to consider when you don't want the topic called, even if the subject is closely related. For example, you might not want that topic called if the user asks "what are the types of cheese you use?" In both cases, the phrases are more information-gathering so I want the user to go to AI Search. I find that I try to come up with alternative phrases beyond "I want..", "I need...", "I want to request.." because (to your point) I would need to put those everywhere. 

I also find that VA configuration is not in isolation. Yes, you have your entities and vocabulary sources in NLU.  Let's say you don't have any of that configured, but you have synonyms in your AI Search profile dictionary, and your catalog items and KB articles have meta tags, descriptive text in the title and body and snippets that answer the user's query, AI Search in Fallback might win.  
Fallback is not bad, it's one of my favorite features of VA - if the KB or catalog content doesn't resolve the users question, the user has the ability to contact a live agent or create a ticket, so I find getting Fallback to be very efficient. 
I hope that helps!