Typo handling and NLU utterance match

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
10-17-2023 02:05 PM
Hi,
I have a Virtual Agent topic related to a single NLU intent. Can anyone help me understand if common typos or misspells are automatically corrected by the NLU or should I be adding them to the Vocabulary for a better match?
Also, a second question on NLU utterances to help understand does a user input need to be a 100% utterance match to trigger the intent or can it be a partial match and still trigger the intent?
Regards,
Ayman
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
10-17-2023 02:09 PM
Hi,
You should not use misspellings in your utterances.
The match does not have to be 100%. You can configure what the confidence threshold (percentage) in the model settings.

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
10-17-2023 02:23 PM
Thanks for your quick response @Raj64
How are misspellings handled in Virtual agent? For e.g. I have an utterance 'I need ServiceNow access' for an Intent that is related to a VA topic. If I typed 'I need access SrviceNow access to perform a task' as part of a Virtual Agent conversation, will this be considered a match and bring up the topic?

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
10-19-2023 06:56 AM
It's a black box. But I too have seen ServiceNow's NLU best practices advising against using misspellings in intents so there is clearly a level of flexibility within the NLU training.
That said, if you already have an intent and utterances, train your NLU Model then click "Test" in NLU Workbench and try out some misspelled utterances and see for yourself what the confidence rating is. From my experience, the only 100% matches will be if your test utterance exactly matches an utterance in the intent (which makes sense) - if you have a typo that's off just a letter or two but otherwise is a good match, I'm sure you'll have a very high confidence rating.
By default, I recall the confidence threshold being 60% - which is very arguably too low - and we've since slowly bumped it up to 80% as we've expanded our model which I'm much happier with.

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
10-20-2023 10:14 AM
Excellent, thanks @Chris D , I guess we just need to find the right balance with the utterances