Train Questions from Real Users

Updated on January 20, 2022

While we try to train the Bot with examples as close to the real world as possible, real user data is still very crucial in the continuous improvement of the Bot.

How is this useful?

After the bot 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 Bot’s current vocabulary, it expands it as well so that the Bot can answer more questions posted by users relevantly in the real world.

How does this work?

The Bot 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 Bot with real user data

  1. Select an intent
  2. 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 bot is  able to answer, and those that the bot is unable to answer.
  3. In the “Confidence” column, it shows how confident the Bot 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:

  1. Add the user question to an existing intent to train the bot
  2. Archive the user question
  3. Correct false positives (when the bot has wrongly identified/predicted the question)
  4. Create a new intent
Did we help you?

Learn more about KeyReply

To see our Conversational AI in action, sign-up for a guided walk-through and see how our platform can be tailored for your organization’s needs