Best Practices for Intents Training

Why should I have best practices while managing Intents? 

Any time a customer calls your business, they do so with a goal in mind. Intents identify these goals by categories and make you more effective. 



A customer might:

  1. Be checking their delivery Estimated Time of Arrival (ETA).
  2. Ask for help with their signup.
  3. Request information on the product.
  4. Report damaged merchandise.


What is an Intent in Talkdesk AI products?

An intent is formed by a set of training phrases, which are examples of different ways of expressing the motivation behind an intent. This motivation (or the training phrases themselves) must remain within the scope of the particular intent it belongs to.

Note: Avoid having similar training phrases with different intents. This will prevent intent collision (an intent being incorrectly detected when the examples or phrases provided are not clear and may be contained on other intents).


If I haven’t configured an intent yet, where should I start? What are the Dos and Dont's of Intent setup?

When customers call your business, list the reasons customers are contacting and prioritize them according to your business needs. Those reasons can become your first intents. You can have more than one intent for each business priority.

  1. Create between three and five intents that are relevant to your business. (For instance: escalation, refund, and order product.)
  2. Separate them into more specific topics (under 50 intents).
  3. Add a minimum of nine training phrases per intent to ensure the model is correctly trained.
  4. Avoid too general or broad intents. (For example, general issues, and general product questions.)
  5. But also don’t be too precise, as creating intents for specific branches or ring groups.
  6. Avoid overlaps between intents by keeping a defined scope and concise training phrases.
  7. For extra accuracy, don’t under train or overtrain the model.


How do I search for good intent training phrases in Interaction Analytics?

  1. Firstly, identify keywords for intents you don’t have training phrases for yet. For the intent “escalate”, keywords could be “agent”, “customer service”, and “let me talk to someone”.
  2. Click on the Search tab in Talkdesk Interaction Analytics™ (IA).
  3. Click enter between each keyword so that they are recognized as a search item.

  1. You can select And to perform the search with all keywords in the phrases, or Or to have at least one keyword in the phrases. Press enter.

5. Then, copy the training phrase and go to Talkdesk AI Trainer™ (AIT). Use the volume arrow icon to inspect the number of occurrences over time.

6. Once the expressions hit the desired volume, create the intent.

7. Then, copy the phrases mentioning the issue and go to Talkdesk AI Trainer™ (AIT).

  1. Choose or create the desired intent and add a training phrase by clicking New phrase.

  1. Edit the phrase, Save, and publish to finish the action.

For more information on training phrases on AI Trainer, please click here.


What are the Dos and Don'ts of the training phrase?

  1. Use phrases:
  • As close as possible to what callers say.
  • Without any interjections or word repetition.
  • As specific as possible to avoid overlapping with other intents.

Example: “Need to book a flight” instead of “Need to book a flight, mine got canceled today” (there’s an overlap with cancellation).

  1. Balance out the number of training phrases across intents.
  2. Avoid:
  • 1-3 word phrases without context.
  • Repetition of the same keyword or similar training phrases in the same intent. It can bias the intent in certain cases or words.
  • Typos and interjections, even if they exist in transcriptions.
  • Long phrases.


How can I reduce bad matches?

  • Identify repeating keywords in an intent, and replace them with synonyms.
  • Delete some mentions of the keywords. 
  • Delete the entire training phrase containing those keywords.
  • Publish the intent with the changes.
  • Wait for new data to see improvements.


Bad matches example:

I need to register.

Alternative: We received an email asking us to sign up.


How can I report mismatches in AI Trainer?

In the Interaction Analytics search tool, click on the detected intent or “Suggest Intent” button next to a phrase.

Choose the correct intent from the intent list.

Go to AI Trainer’s Inbox to see the corrected phrase and use it to train the model. Choose only good quality and diverse phrases to avoid overtraining the model.


How can I manage Intents?

  1. Identify patterns wrongly matched using Interaction Analytics.
  2. Split the intent.
  3. Use the Default Fallback intent. You should add no value and occurring bad matches (as automated messages and menus) here.

One intent can be divided into two or more separate new intents:

  •  Create new intents with different business value.
  •  Create new intents with the phrases you don’t wish to detect in the main intent.


How can I split intents in AI Trainer?

  1. In AI Trainer, split the original training phrases between old and new intents without overlapping training phrases.
  2. Using the search tool in Interaction Analytics, filter by the original intent to find good matches that have a similar pattern for the new separate intents.


Can I increase the AI Trainer confidence threshold per intent?

Yes. Each detection has a confidence threshold score in AI Trainer that improves precision by cutting off unreliable detections.


Start by identifying the confidence threshold, per intent, that allows for the best balance, between good quality and quantity of good matches. After that, select the intent you wish to edit, on AI Trainer, and edit the confidence threshold to the chosen value. Publish the intent and check the results to verify how the improvements affected your performance.


You’ll have:

  • Lower Threshold when there’s:
    • Sensitive detection.
    • More matches.
    • Lower performance.
  • High Threshold
    • Conservative detection.
    • Higher performance.
    • Fewer matches.


While I am adding custom vocabulary, how can I condition the model to correctly recognize expressions that were wrongly transcribed?

In the Models page, select the Speech-to-text model, and click on New phrase.

In the phrase section, write the expression as it is being transcribed, separated by hyphens.

In the Sounds like section, write the expression as it is sounding in the voice recordings, separated by hyphens.

In the Display as section, write the expression in the way you want it to appear, and click on Create.


For more information on Interaction Analytics, please click here.

All Articles ""
Please sign in to submit a request.