What Is Sentiment Analysis? – Definition and How It Works
Learn what sentiment analysis measures, how approaches differ, why aspect-level analysis is more useful, and where accuracy claims break down.
Sentiment analysis classifies the attitude expressed in text — positive, negative, or neutral, sometimes with intensity. It is easy to run and easy to over-trust.
1. The Approaches
| Approach | Basis | Character |
|---|---|---|
| Lexicon | Word lists with polarity scores | Fast, transparent, weak on context |
| Trained classifier | Labelled examples | Domain-specific, good where data exists |
| Language model | Instructed or fine-tuned | Strongest on nuance, costs more per item |
Lexicon methods remain useful precisely because they are inspectable: a wrong result can be traced to a word.
2. Document Sentiment Is Usually Too Coarse
"The camera is excellent but the battery is hopeless" is not positive, negative, or neutral. Scoring the whole review discards the information.
Aspect-based sentiment analysis attributes sentiment to specific subjects — camera positive, battery negative — and that is almost always the more actionable output.
3. What Defeats Naive Analysis
- Negation — "not great" contains a positive word.
- Sarcasm and irony, which invert meaning with no lexical signal.
- Comparison — "better than the old one" is positive about a change, not the product.
- Domain vocabulary — "sick" and "insane" carry different polarity by context.
- Mixed sentiment in one sentence.
- Conditional statements — "would be great if it worked".
4. Accuracy Claims Deserve Scepticism
Reported accuracy is measured against a labelled set, and human annotators frequently disagree with each other on the same text. A model matching human labels 85% of the time is operating near the ceiling of what agreement allows.
That is why an aggregate sentiment score should be read as a rough indicator rather than a measurement.
5. Trends Are More Reliable Than Levels
Absolute sentiment scores are hard to interpret: is 0.62 good? Change over time, measured with a consistent method, is far more meaningful.
Keep the method fixed — changing model or approach mid-series produces an apparent shift that is entirely artefact.
6. Validate on Your Own Data
Off-the-shelf accuracy on general text says little about performance on your domain. Label a few hundred of your own examples and measure.
That exercise routinely reveals systematic errors — a product name read as a negative word, a support phrase misclassified — that are trivially fixable once visible.
7. Where It Genuinely Helps
- Triage, routing negative feedback for attention.
- Trend detection across large volumes.
- Aspect tracking, watching sentiment per feature over time.
- Prioritisation, surfacing what to read rather than replacing reading.
It works best as a filter over volume, not as a substitute for reading what it surfaces.
8. Summary
Sentiment analysis classifies attitude in text using lexicons, trained classifiers, or language models, and document-level scores are usually too coarse to act on.
Prefer aspect-based analysis, treat accuracy claims cautiously given human disagreement, read trends rather than levels with a fixed method, and validate on your own labelled data before trusting any figure.