AI-102 Question 311
Select 2You are building a QnA Maker knowledge base (KB) for a customer service chatbot using Microsoft Azure. After importing a dataset of frequently asked questions, you notice that the chatbot sometimes provides inaccurate answers or fails to respond correctly to similar phrased questions. What should you do to improve the accuracy and relevance of responses from the knowledge base?
- A
Train the knowledge base by adding alternate question phrasing for existing answers.
- B
Enable active learning and review user-submitted feedback to refine the KB.
- C
Increase the threshold for the confidence score of answers in the chatbot configuration.
- D
Use Azure Monitor to analyze the performance metrics of the KB and retrain it as needed.
- E
Manually delete questions and answers that have low user engagement.
Show answer and explanation
Correct answers: A, B
Explanation
To improve the accuracy and relevance of a QnA Maker knowledge base, you need to enhance its training by adding alternate question phrasing and leverage active learning to incorporate user feedback. These actions directly improve the KB's ability to match user questions to the correct answers. Confidence score adjustments and performance monitoring are indirect and insufficient solutions in this scenario.
- A. Correct.
Training the knowledge base with alternate question phrasing improves its ability to match user queries to the correct answers, which increases accuracy.
- B. Correct.
Enabling active learning allows the knowledge base to learn from real-world usage and user feedback, helping refine its responses.
- C. Incorrect.
Increasing the confidence score threshold will filter out low-confidence responses, but it does not directly improve the KB's accuracy or relevance.
- D. Incorrect.
Azure Monitor can provide overall system metrics but does not directly help in training or improving the knowledge base content.
- E. Incorrect.
Deleting questions and answers based on low engagement does not address the core problem of inaccurate responses and may result in loss of useful data.