This tutorial demonstrates how Log Dialogs are employed to better train Conversation Learner models from recorded interactions with real world users.
This tutorial requires that the general tutorial bot is running
npm run tutorial-general
and the weather model created in previous tutorials.
Log Dialogs are recorded logs of your bot's interaction with end users. By harnessing these Log Dialogs you can fix entity labels and action selections to improve the model's performance and overall system performance.
In the web UI, click on "Import Tutorials" and select the model named "Tutorial-11-LogDialogs".
This Model contains one Entity named "city", and Actions designed to respond to inquires about weather in that city. Two Train Dialogs were used to train the Model so performance expectations are somewhat low. The Model would improve with additional training and exposure to real world user interactions.
- On the left panel, click "Log Dialogs", then the "New Log Dialog" button.
- In the chat panel, where it says "Type your message...", type in "Austin weather forecast"
- Click the "Done Testing" button.
- Click the "Austin weather forecast" log dialog from the list.
- Click the "Austin weather forecast" utterance in the chat panel.
- Click "Austin", then click "city" from the Entity List.
- Click the "Submit Changes" button.
- This change in Entity value cause downstream changes to the conversation since we have new entity values in memory. Later actions have likely become invalid especially ones involving the "city" entity.
- Click the "Which city?" utterance in the chat panel.
- Select the response, "The weather in Austin is probably sunny."
- Click the "Save As Train Dialog" button.
- Training is kicked off immediately
One last note. Depending on business needs, the conversation logging feature can be turned off by going to Settings and unchecking “Log Conversations.”
[!div class="nextstepaction"] Entity detection callback