Extracting Sentiment Trends from Topic Miner Clusters

How-To-Articles
1

Extracting Sentiment Trends from Topic Miner Clusters

Executive Summary & Architectural Context

In a contact center processing 10,000 calls a day, Quality Assurance (QA) teams cannot physically listen to enough calls to accurately gauge customer sentiment. While the Genesys Cloud Speech and Text Analytics engine automatically assigns a Sentiment Score (-100 to +100) to every interaction, an aggregate score of “Overall Sentiment: -15” tells you that customers are angry, but it doesn’t tell you why.

The architectural solution is to merge Sentiment Analysis with Topic Miner Clusters. Topic Miner uses unsupervised machine learning to group thousands of raw transcripts into thematic clusters (e.g., “Billing Issues,” “Shipping Delays,” “Password Resets”) without any human pre-configuration. By overlaying the sentiment math on top of these AI-generated clusters, Business Intelligence (BI) teams can pinpoint the exact operational failures driving negative customer experiences.

This masterclass details how to execute a Topic Miner run, interpret the cluster visualization, and extract the underlying sentiment data to build actionable reporting.

Prerequisites, Roles & Licensing

  • Licensing: Genesys Cloud CX 3 (Speech and Text Analytics Add-on).
  • Roles & Permissions: Quality > Topic Miner > View/Edit.
  • Platform Dependencies:
    • A minimum of 30 days of recorded voice or digital transcripts to feed the miner.

The Implementation Deep-Dive

1. Initiating the Topic Miner Job

Topic Miner is not a real-time dashboard; it is an intensive, batch-processed ML job.

  1. Navigate to Admin > Quality > Topic Miner.
  2. Click Create New Mining Session.
  3. Data Selection:
    • Do not mine the entire contact center at once. The results will be too noisy.
    • Filter by a specific Queue (e.g., Retention_Queue) or a specific Wrap-Up Code (e.g., Account_Cancellation).
  4. Timeframe: Select the last 30 to 60 days. The engine needs a massive corpus of text to find statistically significant phrase clusters.
  5. Click Run. This job may take several hours to complete depending on the volume of transcripts.

2. Analyzing the Cluster Map

When the job finishes, Genesys Cloud presents a visual bubble chart. Each bubble represents a “Topic Cluster.”

  1. Size of the Bubble: Represents the volume of interactions containing this cluster of phrases.
  2. Clicking the Bubble: Reveals the most common phrases (e.g., “card declined”, “payment failed”, “late fee”).
  3. The Sentiment Overlay: Look at the right-hand panel. For every cluster, the engine displays the Average Sentiment.
    • If the “Password Reset” cluster has an average sentiment of +45, the IVR and agents are handling it smoothly.
    • If the “Shipping Delay” cluster has an average sentiment of -85, you have found your root cause of customer churn.

3. Promoting Clusters to Active Topics

Topic Miner looks backward. To track these sentiment trends going forward, you must convert the AI clusters into strict rule-based Topics.

  1. Inside the Topic Miner results, select the highly negative “Shipping Delay” cluster.
  2. Click Add to Topic.
  3. Create a new formal Topic named Root_Cause_Shipping_Delays.
  4. The system will automatically map the AI-discovered phrases into the formal Topic definition.
  5. Save and publish the Topic. From this millisecond forward, the Speech Analytics engine will actively tag all future interactions that mention these phrases with the Root_Cause_Shipping_Delays tag.

4. Building the Sentiment Trend Dashboard

Now that the data is actively tagged, you can visualize the trend.

  1. Navigate to Performance > Workspace > Interactions.
  2. Create a custom view.
  3. Filter: Topic = Root_Cause_Shipping_Delays.
  4. Metric: Average Sentiment Score.
  5. Grouping: By Day or Week.
  6. You now have a definitive line chart proving whether the new shipping policy implemented by Operations is actually improving customer sentiment over time, or making it worse.

Validation, Edge Cases & Troubleshooting

Edge Case 1: The “Greeting” Cluster Noise

When you run Topic Miner, the largest, most dominant cluster is almost always generic pleasantries: “Thank you for calling”, “How can I help you”, “Have a great day.”

  • The Trap: These generic phrases skew the overall sentiment highly positive and drown out the actual business intelligence.
  • Solution: You must aggressively prune the Topic Miner results. When you see a bubble containing generic greetings, use the Exclude feature. Exclude these phrases and re-run the miner. This forces the ML algorithm to ignore the boilerplate script and cluster based on the actual body of the conversation.

Edge Case 2: Extreme Sentiment Volatility

Sometimes a specific phrase cluster (e.g., “Cancel my account”) will show a deeply negative customer sentiment, but a highly positive agent sentiment.

  • Troubleshooting: This occurs when an angry customer is screaming, and the agent is forced by QA guidelines to remain excessively cheerful (“I’d be absolutely delighted to help you with that cancellation today!”).
  • Solution: When building sentiment dashboards, do not use the Overall Conversation Sentiment. You must explicitly filter the metric to Customer Sentiment Score. The agent’s forced cheerfulness will mask the customer’s rage if you blend the two scores together.

Official References