Sentiment Analysis

Sentiment Analysis in CX Practice

From Sentiment to Topic-Bound Sentiment

Simple sentiment analysis: "This response is negative."

Topic-bound sentiment analysis: "This response is negative about the topic: Delivery."

The topic-bound approach is far more actionable and is the standard in modern VoC analysis programmes.

Our experience is that the greatest benefit of sentiment analysis is not automation. It is speed. You can identify a growing problem in open responses weeks before it shows in your NPS score. That gives you time to act proactively.

Methods for Sentiment Analysis

Rule-based / Lexical: Based on word lists (positive words, negative words). Fast and transparent, but poor at handling irony and context. Suitable for simple, short responses.

Machine learning-based: Trained on annotated text examples. Good for context-dependent analysis, but requires large amounts of training data and ongoing maintenance.

LLM-based (GPT, Claude, etc.): Uses large language models via prompt engineering. Excellent for nuanced, context-dependent analysis. Scales well and does not require training data.

Typical Use Cases

• Analysis of open NPS responses: What do Detractors vs. Promoters write about?
• Monitoring of Trustpilot/Google reviews: What is the sentiment trend over time?
• Support ticket triage: Prioritize critical inquiries based on negative sentiment
• Social media listening: Identify negative mentions that require a response

Limitations to Be Aware Of

• Irony and sarcasm are frequently misinterpreted
• Short responses (1-3 words) are difficult to classify accurately
• Industry-specific terms may have different sentiment valence than in general texts