Sentiment Analysis — AI Glossary

Reference details

Related terms

• Text Classification
• Bias Detection
• Information Extraction

How it works

1. The label scheme is chosen first and constrains everything: binary polarity, a graded scale, or aspect-level labels attaching sentiment to specific targets within the text.
2. Classical systems fine-tune an encoder for sequence classification, which is cheap, fast and strong when the domain matches the training data.
3. Language models are prompted instead, which removes the labelled-data requirement and handles negation, sarcasm and comparison far better — at a cost per item that matters at volume.
4. Aspect-level extraction identifies the target and its sentiment as a pair, so a single review yields several labelled judgements rather than one aggregate.
5. The document-level aggregate is where information is destroyed. Averaging mixed sentiment produces neutral, which is the same output as genuinely neutral text and means something entirely different.
6. Calibration matters more than accuracy for most uses, since the output is usually consumed as a trend over many items rather than as a judgement about one — and a classifier with a consistent bias still tracks changes correctly.

Example: What the document-level aggregate throws away

Take 1,000 product reviews, of which 300 praise one aspect and criticise another — a common shape: good battery, poor screen. At document level each of those 300 averages toward neutral, so the output is 300 neutral labels that are indistinguishable from 300 genuinely indifferent reviews, and the actionable content of all 600 opinions is gone. Aspect-level extraction on the same 300 reviews yields 300 x 2 = 600 labelled judgements, each attached to a target someone can act on. The reporting difference is stark: document-level says 30% of reviews are neutral, which reads as apathy and prompts nothing, while aspect-level says the screen draws 300 negative mentions, which is a specific defect with a named owner. And the 700 single-aspect reviews are unaffected either way, so the whole gain comes from the 30% that document-level scoring cannot represent. This is why aspect-level is the default for product feedback and document-level survives mainly where the text is short enough to carry one opinion.

Commonly confused with

Sentiment is not intent and not satisfaction, and support systems conflate all three. Sentiment is the emotional tone of the text; intent is what the writer wants done; satisfaction is their evaluation of the outcome. A calmly worded message can describe a severe problem, and an angrily worded one can be a trivial request — so routing on sentiment sends polite emergencies to the slow queue. The related trap is reading sentiment as a measure of the product rather than of the writing: people with mild complaints write more mildly, which is a property of expression rather than of the issue.

When to use it

Reach for it when:

• Trend monitoring over many items, where consistent bias is tolerable and relative change is the signal.
• Aspect-level analysis of product feedback, where the targets are what make the output actionable.
• Triage as a ranking signal among others, informing the order of a queue rather than deciding outcomes.
• Long or nuanced text, where a language model's handling of negation and comparison is worth its cost.

Reach for something else when:

• Document-level scoring of text containing mixed opinions, where averaging to neutral destroys the content.
• As a proxy for urgency or intent, which it does not measure and which route differently.
• Making decisions about individuals — a reply, an escalation, an account action — on a single score with no appeal.
• Across domains without re-evaluation, since sentiment vocabulary is domain-specific and 'sick' is a compliment in some registers.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Sentiment Analysis — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path