Is there any method to train NLU automatically by themselves?

Daichi Ishikawa · ‎05-22-2023

Hi, all! I'm trying to introduce NLU into my VA.

After introducing NLU into the VA, how do I train NLU to adapt the newest environment?

If possible, I would like to achieve this with as little human maintenance as possible.

e.g. NLU trains end-user's feedbacks or new knowledges.

Does anyone have some answer to this question?

Any help would be greatly appreciated!!

WillWitten · ‎05-23-2023

Natural Language Understanding is a great upgrade for the Virtual Agent, so you're definitely on the right track!

When starting to work with NLU, I would advise that you perform 'Intent Discovery' on your tickets (whether they be Incidents, Cases, etc.) and ensure that you properly set the conditions to reduce the noise in your data as much as possible (get good, clean data). Learn more about Intent Discovery on Docs!

Intent Discovery will cluster your data, allowing you to see what your users are currently requesting. You can preview the actual text (usually Short Descriptions), and these should inform you on how your users are wording their requests.

Once you've determined how your users are asking for help, you should build an Intent for each of your Topics, and carefully craft 15-30 utterances for each. Utterances are not keyword phrases - they are natural language, so short sentences, like how you'd ask a person! Be careful not to create conflict within your Utterances for each Intent, otherwise it makes it hard for NLU to resolve which Intent to choose for a given Utterance.

After training your model, run some tests to see how it's working. Once you have it working properly, publish the model and link the topics to their matching Intents.

Once you've deployed your Natural Language Understanding to Production, users will begin asking the Virtual Agent all kinds of questions. When NLU has low confidence on a Prediction, it will add that prediction for review within the Expert Feedback Loop (up to 300/month). This is where your NLU Admin can use real-world utterances to Optimize your models - they will review the Prediction and, if it was incorrect, provide the intent that should have been triggered.

When they have finished reviewing Predictions within the Expert Feedback Loop, the NLU Admin will 'Save Feedback', and these Utterances will be added to the 'Test Set' for a Model, and used to Optimize it. This is how NLU improves over time - through Expert Feedback! Learn more about the Expert Feedback Loop on Docs!

This is how it all works - you carefully create Utterances for each Intent, deploy to Production, then fine-tune the model with real-world results within Production!

I would recommend that any Requests for Knowledge (question-answer type pairs) be left to AI Search - Fallback within the Virtual Agent, and not put into Topics. Enable Genius Cards such as the OOTB Q&A Card to provide an improved experience over the OOTB link output within VA! Learn more about AI Search - Fallback on Docs!

View solution in original post

WillWitten · ‎05-23-2023

Natural Language Understanding is a great upgrade for the Virtual Agent, so you're definitely on the right track!

When starting to work with NLU, I would advise that you perform 'Intent Discovery' on your tickets (whether they be Incidents, Cases, etc.) and ensure that you properly set the conditions to reduce the noise in your data as much as possible (get good, clean data). Learn more about Intent Discovery on Docs!

Intent Discovery will cluster your data, allowing you to see what your users are currently requesting. You can preview the actual text (usually Short Descriptions), and these should inform you on how your users are wording their requests.

Once you've determined how your users are asking for help, you should build an Intent for each of your Topics, and carefully craft 15-30 utterances for each. Utterances are not keyword phrases - they are natural language, so short sentences, like how you'd ask a person! Be careful not to create conflict within your Utterances for each Intent, otherwise it makes it hard for NLU to resolve which Intent to choose for a given Utterance.

After training your model, run some tests to see how it's working. Once you have it working properly, publish the model and link the topics to their matching Intents.

Once you've deployed your Natural Language Understanding to Production, users will begin asking the Virtual Agent all kinds of questions. When NLU has low confidence on a Prediction, it will add that prediction for review within the Expert Feedback Loop (up to 300/month). This is where your NLU Admin can use real-world utterances to Optimize your models - they will review the Prediction and, if it was incorrect, provide the intent that should have been triggered.

When they have finished reviewing Predictions within the Expert Feedback Loop, the NLU Admin will 'Save Feedback', and these Utterances will be added to the 'Test Set' for a Model, and used to Optimize it. This is how NLU improves over time - through Expert Feedback! Learn more about the Expert Feedback Loop on Docs!

This is how it all works - you carefully create Utterances for each Intent, deploy to Production, then fine-tune the model with real-world results within Production!

I would recommend that any Requests for Knowledge (question-answer type pairs) be left to AI Search - Fallback within the Virtual Agent, and not put into Topics. Enable Genius Cards such as the OOTB Q&A Card to provide an improved experience over the OOTB link output within VA! Learn more about AI Search - Fallback on Docs!

Daichi Ishikawa · ‎05-23-2023

@WillWitten

Oh! Thank you for your precious information!! I just needed them!!

I didn't know Expert Feedback, so I'm trying to use it soon.

My VA will be great one through these small step upgrades. So interesting!!