Training Intent Detection Models in AI Trainer

Intent models in Talkdesk AI Trainer are classification models used to automatically uncover the customer’s intentions during calls. They represent the basis for conversational agents, like Talkdesk Virtual Agent, but, moreover, they are the main triggers for any automated task in a contact center.


Why do we need to train intent models?

To improve the automation rate, resolve cases faster, with customer satisfaction, and to know what your end customers are talking about. AI Trainer can be used to observe the intent model performance in production, to train models with real-world data from conversations where intent classification was not satisfactory, and also to easily publish tweaks and improvements to production.

The Intents' page shows the following:


  • The “Observe” section [1] provides information to your contact center staff about model performance in the production environment.
  • The “Train” section [2] is a main work-area, where it is possible to manage intents, training phrases, and entities. The “Inbox” section contains trainable data, collected from live conversations, that could be used to improve the intent model.
  • The” Deploy” section [3] has two functions: to review suggestions for the model improvements, and to manage versions of the model in production.

In AI Trainer, you can manage intents for intent models: create, update or delete intents, as well as add training phrases to intents. Training phrases can be added directly, as made-up or synthetic data, or by labeling real-world data through the Agreement system (explained in the subsection).


To create an Intent, please follow these steps:


1. On the Intents page, click on New Intent [1].


2. Add the “Intent name” [2] and “Confidence Threshold” [3].

3. Add at least a few training phrases each, by clicking New phrase [4].


4. Add “Training phrases” [5] and click Save [6].

Note: The suggested number of training phrases per intent is 10.


5. When you’re done, publish your changes [7].


Entities are metadata associated with an intent. If a speaker mentions specific products, places, or qualities, for instance, that information can be extracted in the context of the intent that is detected. 

With AI Trainer, you can manage custom entities and also annotate training phrases, so they are extracted within the context of an intent.

To create a Custom entity, please do the following:


1. On the Entities page, select New entity [1].


2. Give the new entity a “Name” [2].

3. Add entity values and synonyms [3].

4. Click Create [4] to publish the changes.


For more detailed information on the best practices of how to train an intent model, please consult the full article here.

Filling all the slots

In Virtual Agent, on a conversation, it is usually useful, and sometimes necessary to extract annotated entities, called parameters, when an agent is detecting a contact’s intent. If the parameter is required but not detected, then Virtual Agent will ask for this information - and fill in all the slots in the conversation.

In AI Trainer it is possible to configure parameters, automate prompts for when a parameter is not detected, as well as determine automatic responses for a detected intent.


To configure a parameter, please do the following:


1. Open the training phrase [1] containing words to be annotated.

2. Highlight words [2] and a list of entities will appear [3].

3. Save [4] the training phrase and the created parameters.


4. Open the list of parameters [5], and select one to edit [6].

5. Tick the option “If enabled, allows configuring prompt responses” [7] to add an automatic prompt, and to fill in the slot defined by the parameter.

6. At the end, select Save [8].


Managing the agreement system

AI Trainer ensures that agents, as domain experts, can label data without decreasing the model’s performance by accidentally deploying the changes to production.

 This is possible through the Agreement process, which happens in stages:

  • Multiple agents can independently make suggestions for a model change.
  • Supervisors can review these suggestions and decide to move on with the change or decline it.
  • Post review, only supervisors, can deploy models to production.


In the Inbox section, Agents can see all the data that can be labelled. Each Agent has their own Inbox. If Agent A thinks a phrase should be labeled as "student loan", Agent B can think otherwise. This is what defines the overall agreement percentage (%), when the phrase moves to quality check. 

1. Agents can select one phrase to label [1], and either ignore it [2], flag it for later [3] (To-do later) or suggest this phrase as a training phrase for intents [4] (Train intent).


On the Quality check page, supervisors can see all agents’ suggestions. To review them, please follow these steps:


1. Click on the suggestion [1].


2. Click on the “Start review” button.

3. Approve each suggestion based on the level of agreement between agents [2], and go to the Launch section [3].


4. Click on the reviewed suggestions [4] and on the settings menu [5] to choose whether to launch the changes right away or schedule the deployment in the future [6].

When the model is trained, check the Intents' page to confirm if the new training phrases are properly added.

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