Sentiment Analysis
Deciding whether text is positive or negative — the most deployed NLP task, and the one whose target may not exist.
When not to use it
- On individual high-stakes decisions. It's an aggregate instrument. Routing one customer on it is misusing the tool.
- When you needed aspect-level detail. "Negative" tells you nothing about what to fix. Aspect-based sentiment does.
- Across domains without checking. "Predictable" is negative for films and positive for delivery. The model only knows what it saw.
- On sarcasm-heavy text. It's unsolved, plausibly unsolvable from text alone, and humans aren't good at it either.
Reach for something else instead
- Aspect-based sentiment — sentiment per topic. Usually the thing you actually wanted.
- Direct measurement — churn, returns, NPS. If you can measure the behaviour, don't infer the feeling.
- Lexicon methods — instant, free, interpretable, and a fair baseline on social text.
- Emotion or stance classification — richer, with their own construct problems.
Sources & further reading
- Pang & Lee (2008), Opinion Mining and Sentiment Analysis — the founding survey; still clear about what the task is and isn't.
- Socher et al. (2013), Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank — the benchmark, and where compositional sentiment got taken seriously.
- Hutto & Gilbert (2014), VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text — the lexicon baseline that keeps being competitive.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Treating a document-level score as actionable. It aggregates away the information you needed.
- Ignoring domain shift. A movie-review model on support tickets is measuring the wrong vocabulary.
- Reporting accuracy above the annotator agreement rate without noticing what that implies about the labels.
- Assuming sentiment is a property of text. It's a property of a reading, and readings differ.
- Using it on individuals rather than trends. It's a thermometer for a population, not a diagnosis for a person.
At a glance
FieldApplied AI
Worksaggregate trends on clear text
Failsindividuals, sarcasm, mixed opinions, domain shift
Real ceilingannotator agreement, often 70–80%
The upgradeaspect-based sentiment
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Document-level vs. aspect-based sentiment — one number for the whole review vs. one per thing discussed. The first is easier to dashboard; the second is what you can act on.