Are there always exactly six basic emotions?

Different theories propose different sets.

Does no applicable label mean no emotion exists?

No; it describes the annotation outcome.

AI Glossary / Natural language and speech / Advanced

Basic emotions

Definition

Basic emotions are a limited set of emotion categories proposed as fundamental within a particular theory of emotion. In natural-language processing, such categories often provide labels for describing emotion expressed in text. The schemes differ, and choosing one for a dataset does not establish that it is a complete or universally agreed description of human emotional experience.

What it means

A categorical representation gives an emotion a name, such as anger or sadness. This differs from a dimensional representation that describes affect using coordinates such as pleasantness and arousal.
The word basic does not mean trivial, weak or easy to detect. It describes the proposed status of a category within a theory. In an AI project, it is essential to separate that theoretical claim from an operational labeling decision. A model trained to assign six labels learns a task defined by those labels; its output does not independently validate the theory behind them.

How it works

A project selects a category scheme and defines what annotators should label. Possibilities include emotion explicitly expressed by a writer, emotion attributed to a character, or emotion evoked in a reader. These are different targets and should not be silently combined.
The labeling instructions specify whether more than one category is allowed, how uncertainty is handled and what to do when none of the categories applies. A classifier can then be trained and evaluated against annotations produced under those instructions. Agreement with annotations measures performance on that task, rather than direct access to someone's private experience.
Ekman's commonly cited six are anger, disgust, fear, happiness, sadness and surprise. Plutchik's scheme uses eight, including trust and anticipation. Alternative dimensional and appraisal approaches represent emotion differently. Naming the chosen scheme makes a model's outputs easier to interpret and prevents incompatible category sets from being mistaken for equivalent measurements.

Three practical examples

A research team labels fictional dialogue for explicitly expressed emotion. It allows both fear and sadness on the same passage, preserving mixed expressions instead of forcing a single category.
An editor compares how anger and happiness are expressed by characters across a collection of stories. The analysis concerns textual portrayals, not the emotions of the authors or readers.
A customer-feedback project studies emotion language associated with reported service problems. Reviewers examine the underlying messages and issue descriptions rather than treating a predicted emotion as a verified personal state.

Think of it this way

Think of a set of labeled drawers used to organize descriptions. The drawers make comparison easier, but their labels and boundaries were chosen for a purpose. A description may fit more than one drawer, or none. Sorting it does not prove that every human experience comes in exactly those forms.

Worked example

Consider five invented messages and a labeling task that records explicitly expressed emotion using Ekman's six-category vocabulary. Multiple labels are allowed. The task makes no claim to verify how the fictional writers actually feel.
Message A says, “I am happy that the project succeeded.” Its assigned category is happiness. Message B says, “I am frightened about tomorrow and sad about leaving.” Its assigned categories are fear and sadness. Message C says, “I am angry about the decision and sad about the outcome.” It receives anger and sadness. Message D says, “I am surprised by the announcement.” It receives surprise. Message E says, “The meeting starts at ten.” It receives no emotion category under these instructions.
Across five messages, sadness appears twice: 2/5 = 40%. Happiness, fear, anger and surprise each appear once: 1/5 = 20% each. Disgust appears zero times. The category percentages sum to 40% + 20% + 20% + 20% + 20% = 120%.
That total is valid because this is a multilabel task. Six category assignments have been made across five messages, and two messages each carry two labels. The denominator is messages, not assignments. If the question instead asks what fraction of the six assignments is sadness, the answer is 2/6, or about 33.3%. That is a different statistic.
The unlabeled fifth message is not evidence that its writer has no emotions. It means the text contains no expression meeting this task's instructions. Similarly, a zero disgust count concerns these five annotations; it does not establish that disgust is absent from a wider population.

Illustrative case study

In a fictional dashboard project, the worked percentages trigger an error because the reporting template expects category shares to total 100%. A developer proposes normalizing the six assignments, turning sadness's message prevalence from 40% into an assignment share of roughly 33.3%.
The project lead recognizes that the proposed change answers a different question. The dashboard is revised to state “percentage of messages carrying each label” and to explain that labels can overlap. A separate count records messages with no applicable emotion label.
The team also changes the display from “customer emotional state” to “emotion expressed in annotated text.” This makes the measured target explicit. It preserves access to the original message for interpretation and avoids turning a limited annotation exercise into a claim about a person's inner condition.

Common misconceptions and mistakes

There is no single category inventory that every emotion theory accepts. Six-category and eight-category outputs cannot be compared by assuming their labels and boundaries are identical.
Emotion classification is also not merely positive-versus-negative sentiment classification. Surprise, for example, does not specify whether an event is welcome. A word associated with anger can appear in a quotation, a denial or a discussion of another person's reaction. Context and the annotation target determine how it should be handled.

Business use

Document the category scheme, target of annotation and permitted label combinations before interpreting model reports. Distinguish message prevalence from assignment shares and model scores from verified feelings. Use the analysis to investigate communication patterns and stated problems, with evaluation appropriate to the language and context being studied.

Check your understanding

Why do the example percentages total 120%?

Answer: Categories overlap across messages.

What fraction of messages expresses sadness under the instructions?

Answer: Two of five, or 40%.

What changes if the denominator is six assignments?

Answer: Sadness becomes about 33.3% of assignments.

Related terms

Sentiment analysis
Affective computing
emotion classification
valence
arousal
annotation
multilabel classification
inter-annotator agreement

Learn next

Compare categorical and dimensional emotion representations, then examine how annotation instructions influence evaluation results.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 23, section 23.1: categorical and dimensional theories of emotion. https://web.stanford.edu/~jurafsky/slp3/23.pdf W3C Working Group Note, Vocabularies for EmotionML: alternative vocabularies and Ekman's six-category inventory. https://www.w3.org/TR/emotion-voc/ The five messages, calculations and fictional dashboard case are original teaching illustrations.

Previous: Baseline Next: Basis