To achieve good classification accuracy, it’s important to provide your agent with enough data. The greater the number of natural language examples in the User Says section of Intents, the better the classification accuracy. We encourage you to use Example Mode for your User Says examples instead of Template Mode, since the former provides better data for machine learning.
When you create a new intent, start with examples that have the most number parameters. This way you will define what entities should be used in this intent and name all the parameters the right way. Having annotated the first few long examples, it will be easier for you to continue with shorter ones, as the system will start suggesting the correct entities for new examples.
To make the training process more efficient, we have created a Training tool that allows you to analyze conversation logs with your agent and add annotated examples to relevant intents in bulk.
How It Works
As you and your users chat with your agent, you can access the conversation logs by clicking Training in the left side menu.
The logs are presented in two views:
- Training - This view shows conversations with the agent for review and performance improvements. Each user request is a list item, showing the intent that will be used for processing, as well as the current parameter annotation. You can reassign inputs to correct intents and fix annotations. Every time you approve changes, the agent is trained, and the results in the tab are updated.
- History - This view displays the conversations in a plain mode. This way you can see latest conversations with your agent in chronological order.
You can upload sample user inputs in a
.txt file or in a
.zip archive with multiple (up to 10)
.txt files. Each input should start from a new line.
Just click the Upload button in the right upper corner.
How to Train Your Agent
Click on a dialog (dialogs are named by the first user input in the session). You may see that some inputs don’t match to any intent or have incorrect annotations.
Handle Unmatched Inputs
Unmatched inputs are marked by an exclamation mark error_outline. You can assign unmatched inputs to intents in two ways:
- Add inputs to one of the existing intents
- Create a new intent with this input
In the case of incomplete or incorrect annotations, you can fix it the same way as adding or editing examples in intents.
To add an annotation, highlight the word that should be annotated and select an entity from the list.
To edit an existing annotation, click on an annotated word and select a different entity from the list.
If an entity used for annotation doesn’t contain the annotated word/phrase, it’ll be automatically added to this entity as a new entry unless the 'Allow automated expansion' option is checked.
Add Examples to Intents
When you edit an input, a green check mark check appears on the right. It means that the input will be added to the assigned intent.
If you don’t want to add the input, click on the cancel icon not_interested right below the check button.
Once you’ve reviewed all the inputs in the session, click the Approve button to add all the fixed inputs to the respective intents. Your agent will start training immediately. It’ll show a notification when the training has been completed.