Glossary
Sentiment vs theme extraction
Sentiment extraction labels each statement as positive, negative, or neutral; theme extraction labels each statement by what it is about. For organisational decisions, theme extraction is the load-bearing signal — sentiment is a low-resolution proxy that hides which structural pattern is producing the feeling.
Sentiment extraction and theme extraction are two distinct text-analysis tasks. Sentiment classification answers "How does this person feel?" with a polarity label. Theme (or topic) extraction answers "What is this person talking about?" with a categorical label drawn from a code book.
For employee feedback, sentiment scores tend to converge to a narrow band — most comments are mildly negative — and offer leaders little to act on. Theme extraction, by contrast, lets leaders see that 38% of the negative comments are about decision bottlenecks, 22% about approval loops, and so on. The theme tells you which lever to pull; sentiment only tells you that something is wrong.
The Frustration Condition Tracker uses theme extraction (against the five Frustration Architectures) as the primary signal. Sentiment is retained as a secondary, low-resolution check, not as the headline metric.
How it is measured
- Theme accuracy: per-Architecture precision and recall against a held-out coded sample. Target ≥ 0.8 per major class.
- Sentiment accuracy is reported but not used as the headline KPI; the headline is the count of statements per Architecture cluster.
Dr. Tim Hough · ISBN 979-8-9965397-1-0 · Buy the book →