Professional Documents
Culture Documents
Eric Wayne
STSM, IBM Watson AI
Assistant Skills
Dialog Agent
1. The Practices
2. The Tools
3. The Story (an example)
Cost
Revenue
Engagement
“If your goal is to reduce customer service cost, it should cost less to maintain your
assistant than it does to staff the equivalent customer support team capacity…”
Think 2019 / 4543 / Feb 2019 / © 2019 IBM Corporation
Establishing a Baseline – Process Criteria
Measure
Live System
Analyze
Deploy
Improve
Pre-deploy
Testing
Coverage is a view of the range and depth of subject matter your assistant is trained on
Coverage can be measured live in production and offline with test sets or historic logs
• Conversation containment
• Conversation success (task completion)
• Intent confidence of messages in the conversation
• Precision of individual messages in the conversation
• Sentiment analysis
• Explicit user feedback (NPS at end of a sample or all conversations)
Think 2019 / 4543 / Feb 2019 / © 2019 IBM Corporation
Measure Coverage and Effectiveness
(1) Measure (2) Sample & (3) Analyze (4) Update workspace
Label
See what users are saying, and make updates to intent and entity training in your development
workspace.
Recommenders for conflict resolution and entity expansion in
Watson Assistant
Jupyter notebooks that help customers implement the best practices – beginning with
Measure and Analyze Effectiveness. Available in Github and the Watson Studio community.
The Story (an example)
Assistant Skills
Dialog Agent
Measure
Bootstrap
Live
System
Analyze
Evaluate
Deploy
Bootstrap
Improve
Pre-deploy
Testing
Jupyter notebooks that help customers implement the best practices – beginning with
Measure and Analyze Effectiveness. Available in Github and the Watson Studio community.
Watson Assistant Continuous Improvement
Measure
Bootstrap
Live
System
Analyze
Evaluate
Deploy
Bootstrap
Improve
Pre-deploy
Testing
Measure Notebook retrieves logs and computes
automated metrics
Measure – Coverage over time
Measure – Effectiveness detail
Export an assessment spreadsheet. Let’s focus on
conversations that were escalated to an agent
Watson Assistant Continuous Improvement
Measure
Bootstrap
Live
System
Analyze
Evaluate
Deploy
Bootstrap
Improve
Pre-deploy
Testing
In this scenario, we are focusing just on conversations that escalated
when a message was not covered by the chatbot
Measure – Annotate a sampling of responses
Open the Analyze Effectiveness Notebook and load the
annotation spreadsheet
Analyze Effectiveness - Summary Metrics
Dialog
• Prioritize implementing the missing dialogs
based on which cases lead to escalations
Assistant Skills
The provision of the information contained herein is not intended to, and
does not, grant any right or license under any IBM patents, copyrights,
trademarks or other intellectual property right.
Eric Wayne
STSM and Development Manager, IBM Watson
—
ewayne@us.ibm.com
ibm.com