Can one sentence contain several sentiments?

Yes, toward different targets or aspects, and sometimes conflicting evaluations of the same target.

Does positive language always mean support for a topic?

No. Target, attribution, comparison, and the task's definition of stance matter.

Can a factual-looking statement convey sentiment?

It can imply evaluation in context, but the interpretation should not be assumed from a word list alone.

AI Glossary / Natural language and speech / Advanced

Sentiment

Definition

In natural language processing, sentiment is an expressed or implied evaluation, commonly described as positive or negative toward a target. Depending on the task, annotations may also represent neutral, mixed, or more detailed degrees of evaluation. Sentiment concerns evaluative meaning in context; it is not automatically a measurement of a person's internal emotional state.

What it means

A customer can praise a product's instructions and criticize its construction in the same sentence. A single positive-or-negative label may hide that contrast. Identifying the target and, when relevant, its aspect helps preserve what the evaluation is about.
Sentiment is related to emotion but is not identical to it. A calm assessment can be unfavorable, while an emotional statement need not evaluate the product being studied. The words themselves also need context: a negative-looking expression may be quoted, negated, or directed at a different target. An annotation policy should state whose evaluation is being recorded and what counts as neutral or mixed.

How it works

Sentiment analysis assigns a representation to evaluative language. A lexicon-based method starts with word or phrase scores and applies rules; a learned model predicts labels or scores from examples. Both approaches need a clearly defined unit of analysis, such as a document, sentence, or target - aspect pair.
Context can alter interpretation through negation, comparison, contrast, sarcasm, and attribution. A system must connect the evaluation to the appropriate target rather than simply count favorable and unfavorable words.
A polarity score, a sentiment-intensity rating, and a model's confidence are different quantities. A high probability of a negative label does not necessarily mean strong negativity. The meaning of every score should be documented before it is aggregated or compared.

Three practical examples

A product team separates praise for setup instructions from criticism of durability. This preserves actionable distinctions that one overall review label would lose.
A service dashboard distinguishes the writer's complaint from praise quoted from an advertisement. Attribution prevents the quoted claim from being mistaken for the customer's own assessment.
A content analyst compares evaluations of two competing services within the same post. The same sentence can favor one target and disfavor the other, so a target-free label may be misleading.

Think of it this way

Think of a review as a set of labeled arrows: favorable or unfavorable, each pointing toward something. Counting arrows without checking their destinations can produce the wrong conclusion. The arrow's direction, its strength, and your confidence that you interpreted it correctly are three separate matters.

Worked example

Use three invented reviews and two aspects: instructions and housing. Review A says, The instructions are clear, but the housing is flimsy. Review B says, The instructions are clear and the housing is sturdy. Review C says, The instructions are confusing, but the housing is sturdy.
Under a declared toy policy, clear and sturdy are positive evaluations of their respective aspects; confusing and flimsy are negative. The six aspect evaluations contain four positives and two negatives. Instructions has two positives and one negative; housing also has two positives and one negative. Each aspect therefore has a positive share of 2/3, or 66.7%, in this tiny supplied sample.
Assigning +1 to positive and −1 to negative gives an overall mean of (4 − 2)/6 = 1/3. That number does not mean every reviewer is mildly positive. Two reviews contain opposing aspect evaluations, while one is positive on both aspects. A classifier that labels every aspect positive gets 4/6 = 66.7% accuracy yet misses both negative evaluations. These counts demonstrate aggregation and evaluation pitfalls; they do not estimate a real customer population.

Illustrative case study

Fictional manufacturer PlainBuild receives comments about a new storage unit. Its first dashboard summarizes each review with one label. Editors notice that praise for clear assembly instructions can coexist with concern about the housing.
They introduce separate aspect annotations and retain the supporting phrases. The reporting team shows positive, negative, mixed, and unclassified cases according to a written policy, rather than treating uncertainty as neutrality. They also examine whether the collected comments represent the audience they intend to understand.
This fictional case illustrates a more interpretable reporting design. It offers no claim that sentiment labels reveal all customers' feelings or establish the cause of a product problem.

Common misconceptions and mistakes

Sentiment is not the same as truth. A confidently expressed positive review can contain an inaccurate claim, and a negative review does not independently prove a defect.
Neutral, mixed, and unknown should not be silently merged. A statement with no evaluation differs from one that contains both praise and criticism, and both differ from an unresolved interpretation. Exact labels depend on the task policy.
Do not transfer word scores mechanically between domains or ignore negation and attribution. Avoid interpreting a confidence score as intensity. Finally, a large collection of posts is not automatically representative of all users; repeated contributions and collection choices affect what an aggregate describes.

Business use

Sentiment can help organize feedback, identify recurring aspect-level concerns, and support review of communication patterns. Keep the target, evidence text, date, and interpretation policy available for inspection. Evaluate errors by label and by relevant language variety or domain, rather than relying only on overall accuracy. Connect findings to follow-up investigation; an evaluative pattern is useful evidence about the collected messages, not a complete explanation of customer behavior.

Check your understanding

Why are Reviews A and C mixed in the exercise?

Answer: Each combines a positive evaluation of one aspect with a negative evaluation of the other.

Why is 66.7% accuracy insufficient for the always-positive classifier?

Answer: It misses every negative aspect evaluation in the sample.

What does the mean of 1/3 hide?

Answer: The distribution across aspects and reviewers, including the two mixed reviews.

Related terms

Sentiment analysis
Connotations
Stance detection
Negation
Subjectivity
polarity
opinion mining
aspect-based sentiment analysis
emotion
connotation

Learn next

Write an annotation guide for a small feedback sample, specifying targets, aspects, attribution, and treatment of mixed or uncertain cases. Compare aspect-level results with one label per review, then inspect which information is lost in aggregation.

Sources

Jurafsky and Martin, Speech and Language Processing, Chapter 5, section 5.1, sentiment and affective meaning: https://web.stanford.edu/~jurafsky/slp3/5.pdf Liu, Sentiment Analysis and Opinion Mining, 2012, Chapter 2, opinion targets, subjectivity, and emotion: https://www.cs.uic.edu/~liub/FBS/SentimentAnalysis-and-OpinionMining.pdf The reviews, arithmetic, and PlainBuild scenario are original illustrations.

Previous: Sentence separation Next: Sentiment analysis