VA conversation ends when typing certain words/phrases

[Yasmin]

When Typing certain phrases or acronyms (e.g.: "hr"), VA triggers the Closing Conversation in Custom Greetings and Setup. This topic is not using NLU to invoke it because the Closing Conversation topic does not have any intent associated to it.

 

A second problem we are seeing with this is that it also invokes the Anything Else topic right after the Closing topic. 

Why would a word/acronym/phrase trigger the conversation to end? How can we avoid the session from ending? 

A 3rd problem we are runing into is that if I type "HR" in uppercase it triggers the correct topic (live agent). Whereas if I type "hr" in lowercase it automatically ENDS the session.

Note: This is happening only in the live environment. The behavior in subprod environments differs, it is more accurate, even though the setup topics and the NLU model are a copy of the subprod environment. 

[Brian Bakker]

Hello Yasmin,

Have you defined HR as a Vocabulary Item using the regex "\b(?i)hr\b" in your NLU Model with the synonym "Human Resources"?

Vocabulary Item List View

Once added, please retrain the NLU Model and see if that resolves this issue.

Best regards,

Brian

[Yasmin]

Thanks Brian, I was under the impression that as of Paris NLU was case unsensitive per this previous post. So I guess my question is why do we have to define "hr" as a synonym if it's just the lowercase form of "HR". It works for other words, but for some reason in this scenario it triggered the function to end the VA session.

I added "hr" as a synonym, and it is invoking the correct topic now. 

[Brian Bakker]

Hello Yasmin,

Please review the following articles -

NLU Service updates [San Diego]

• Utterances for all languages have been made case insensitive during intent prediction. If you want to keep certain words or phrases as case sensitive, you can create pattern vocabulary using regular expressions.

KB0995575: [VA/NLU] Vocabulary (Dictionary) in the utterance is not recognized and it returns an incorrect intent and the wrong topic is displayed in Virtual Agent or it returns no intents above the confidence threshold

• However, NLU has always been case sensitive, so we have built in tolerance in the neural model to mitigate this issue in Quebec, but there are some words where the neural network is biased e.g "HI" (Hawaii ) is seen so many times in its corpus, that is is biased towards it being a noun. These are the places where our build in tolerance fails and we expect users to add these samples to their NLU Model to override the ServiceNow NLU Service and its general corpus.

• Therefore the Quebec NLU Service works for most vocabulary, however, there are still some outlier cases where it is still case sensitive or if there are specific words that needs to be case specific, use regex-based vocabulary so that the case specific meaning is used to understand the intent.

Therefore, as "HR" is an acronym, you must define the synonym "Human Resources" as a Vocabulary Item for the vocabulary "HR". If the predictions are inconsistent between the uppercase and lowercase versions of the vocabulary in your test utterances, a regex expression for this Vocabulary Item can fix this behaviour.

I hope this helps.

Regards,

Brian