While we try to train the virtual assistant with examples as close to the real world as possible, real user data is still very crucial in the continuous improvement of the AI-generated assistant.
How is this useful?
After the virtual assistant is live, it is important to view the data collected from questions users ask to continuously train the language model. Adding user questions for training not only improves the virtual assistant’s current vocabulary, it expands it as well so that the virtual assistant can answer more questions posted by users relevantly in the real world.
How does this work?
The virtual assistant automatically picks out real user data and parks it under the most relevant intent. For every piece of data received, you can decide whether to accept it as an example for the current intent, archive it, classify it under a different intent, or create a new intent altogether.
Steps to improve the virtual assistant with real user data
- Select an intent
- Navigate to the ‘Chat Log’ tab to view the list of “Unclassified Messages”. These are all real user questions the Bot has automatically picked to fit under this intent. This includes both questions that the virtual assistant is able to answer, and those that it is unable to answer.
- In the “Confidence” column, it shows how confident the virtual assistant is about the example and intent match it had done, to the max limit of 1.0. The higher the confidence level, the better.
From here there are four actions you can take: