---
name: NLP & Sentiment Analysis
description: Turn free text — reviews, social posts, support tickets, survey open-ends — into structured signal with natural-language processing: themes, sentiment, and emerging issues, reported with confidence and grounded in real quotes rather than a single averaged mood score.
audience: growth marketer · product manager · researcher · support lead
---

# NLP & Sentiment Analysis

## What this is
A method for analysing large volumes of free text: extracting themes and entities, scoring sentiment, spotting emerging complaints or demand, and tracing every finding back to the actual sentences that support it.

## What this is NOT
- **Not a mood oracle.** Sentiment scores are approximate; sarcasm, mixed opinions, negation, and domain slang all fool them. A single "72% positive" number is reported with its error bars and its caveats, never as fact.
- **Not surveillance or manipulation.** It analyses consented/public feedback in aggregate; it does not profile named individuals or engineer persuasion from private data.
- **Not a substitute for reading.** The best insight often comes from the quotes it surfaces, which a human reads — the model routes attention, it does not replace judgment.

## When to use
Making sense of thousands of reviews or tickets; tracking sentiment over time or across releases; detecting an emerging issue before it trends; coding survey open-ends at scale.

## Operating principle
Quotes over scores. Every theme and sentiment claim is anchored to representative verbatim text, so a reader can verify it — the number is a pointer to evidence, not the evidence.

## Capabilities
- **Theme & entity extraction** — Goal: what people are talking about. Method: topic modelling / clustering + entity and aspect extraction (which feature, which sentiment), dedup near-identical posts, report volume per theme. Output: a ranked theme list, each with example quotes and a share-of-voice count. Quality bar: themes are backed by real quotes; volume counts exclude duplicates and bots.
- **Sentiment & aspect scoring** — Goal: how they feel, about what. Method: aspect-based sentiment (not one blob), model confidence reported, sarcasm/negation flagged as a known limit, validated against a human-labelled sample. Output: sentiment by aspect over time + an accuracy note from the labelled check. Quality bar: the score's agreement with human labels is stated; low-confidence text is flagged, not force-scored.
- **Emerging-issue detection** — Goal: catch it early. Method: track theme velocity and anomaly spikes, separate signal from seasonal noise, alert on a rising complaint with its first appearances. Output: an early-warning list with the originating quotes. Quality bar: an alert always shows the underlying text so a human confirms it's real before it escalates.

## A worked example
"What are people saying since the redesign?" → Themes ranked; "checkout confusion" is up 3× in two weeks (aspect-sentiment sharply negative), shown with eight verbatim complaints and their dates; overall "positive %" reported with a note that the labelled-sample accuracy is 0.82 and sarcastic reviews are under-counted. The team reads the eight quotes, not just the bar.

## Guardrails & escalation
Named-individual profiling or private-data analysis → declined; route to legal/privacy. Crisis-level negative spikes → the Crisis PR skill. High-stakes accuracy (safety, medical, legal complaints) → human review of every flagged item. Aggregate and consented only.

## References
Standard NLP practice (topic modelling, aspect-based sentiment, NER); inter-annotator agreement for validation; GDPR Art. 5–6 for personal data; platform terms of service for collection. Verify model accuracy on a domain-labelled sample.