Article Quick Links:
What is Bot Training?
Bots understand chatters' questions by comparing them to a list of sample questions. We call these sample questions the Training Questions. Each Answer has its own unique list of Training Questions.
The bot does more than find exact matches or keywords when comparing Answer Training—it uses an algorithm to analyze the answer training as a whole, looking for patterns. Things like synonyms and phrasing variations are all accounted for.
You bot can do all this in multiple languages, too. If your bot is equipped with the multilingual feature, you can train your bot in any language that you've enable in your multilingual settings.
Answer Training is located at the top of the workspace of the Answers view on the bot dashboard.
How Bots Use Training
When a chatter asks a Question, the bot references its Training before it decides how to answer. The following occurs:
The bot compares the chatter’s Question to the Training Questions
If the bot is confident that it has a match, it will provide that Answer
If the bot is not confident, it will provide a Not Understood or Needs Clarification Answer.
As the bot’s understanding of a particular Answer is reflected in the Answer's Training. The larger the sample of Training Questions, the higher likelihood that the bot will understand the Question. The Ada Machine Learning Team recommends the following best practices:
- The Training indicator bar (next to the Training heading in the Answer editor) changes colour from red, to yellow, to bright green as the number of Training Questions approaches 10. This bar only reflects the quantity, not the quality of training, so ensure all of your Training Questions are relevant.
- New Answers should start with at least 7-10 training questions for each new trained answer.
- Over time, aim to have training balanced across answers. Bots perform better when answers have a similar number of training questions. Approximately 20 Training Questions per answer is recommended.
Training Best Practices
The Training process requires an initial proactive period, followed by ongoing management. However, as the bot learns how to respond to chatters’ questions, adding new Training Questions isn’t as important as maintaining and reviewing the Training Questions in place.
The bot’s Training should:
- Expand your training Questions to consist of more than questions – they should include a combination of queries, phrases, and keywords.
- Use synonyms and creative sentence composition to prepare your bot for various chatter questions
- As you create Training Questions, consider different perspectives and levels of experience with your product/service
- When an Answer’s Training is limited, the bot will be less confident about how to respond and will be more likely to respond with Not Understood. To avoid this, we encourage you to provide thorough, high-quality Training for each Answer.
Be Specific + Unique
- Training Questions should be directly related to the Answer content.
- A question cannot be trained for multiple Answers. If you do attempt to add a duplicate question, you will be provided with messaging identifying where the duplicate exists. None of the questions in the batch will be trained until the duplicate is removed.
- If multiple Answers have similar Training, there are two solutions:
1) Create a general Menu Answer with Training that accounts for all the specific sub-Answers (move Training from specific Answers to Menu Answer). Direct chatters to specific Answers using Quick Reply buttons in the Menu.
2) Merge the Answers into one and consolidate the Training.
Stay Up to Date
- Training requires ongoing review and optimization. We recommend that you regularly audit Training Questions for each Answer and remove Training that is irrelevant. For example, remove Training that includes personal situations, scenario-based questions, questions in languages other than the bot’s native language, or phrases that are unrelated to the bot’s content.
To learn more about how to train your bot, see this article.
Have any questions? Contact your Ada team—or email us at firstname.lastname@example.org.