How similar are “joy” and “surprise”? Conversely, how far apart are “joy” and “sadness”?
We naturally organize words that express emotions—such as “happy,” “afraid,” and “angry”—according to similarities and differences in their meanings. However, investigating this entire structure using human participants alone is not easy, because the number of comparisons rises rapidly as more emotion terms are included.
Specially Appointed Assistant Professor Ke Han and Associate Professor Eiji Watanabe of the Laboratory of Neurophysiology, National Institute for Basic Biology (NIBB) investigated the semantic organization of emotion terms by asking GPT-4 to arrange those terms in a spatial layout and comparing the results with human judgments.
The results showed that prompting GPT-4 with a task in ordinary language reproduced a structure closely resembling the arrangements made by human participants. Moreover, when the vocabulary was expanded to 99 terms, a structure related to emotional intensity, or arousal, emerged that had been difficult to detect using only a small number of emotion terms. These research findings were published in Scientific Reports .
[Background]
Emotions can be represented not only as “categories,” but also as “positions”
Emotions can be divided into categories such as joy, anger, and sadness. In psychology, however, emotions can also be represented as positions along several continuous dimensions. Three representative dimensions are:
Pleasure: whether an emotion feels pleasant or unpleasant
Arousal: whether an emotion is excited and activated or calm and subdued
Dominance: whether a person feels in control of the situation or overwhelmed by it
For example, “excitement” and “serenity” are both relatively pleasant emotions, but they differ greatly in arousal. Likewise, “rage” and “sadness” are both unpleasant emotions, yet they also differ in arousal.
Such emotional structures have traditionally been studied mainly by asking people to rate how similar emotion terms are to one another. However, experiments that comprehensively compare large numbers of emotion terms place a very heavy burden on participants.
[Research Findings]
Placing three emotion terms on a “map”
The research team developed a task in which three emotion terms were placed simultaneously on a grid displayed on a screen. For example, when the words “joy,” “surprise,” and “sadness” were presented, terms judged to be similar in meaning were placed close together, whereas terms judged to be different were placed farther apart. Placing three terms at once made it possible to express three relationships—joy–surprise, joy–sadness, and surprise–sadness—in a single trial.
In the first study, 89 human participants and GPT-4 performed the same task using six basic emotion terms. Their results were also compared with those obtained using conventional word embeddings, which represent the meanings of words as numerical vectors.
GPT-4 produced arrangements similar to those made by humans
Analysis of the six basic emotions showed that both humans and GPT-4 formed the same two groups:
“Joy” and “surprise”
“Anger,” “fear,” “disgust,” and “sadness”
By contrast, the word-embedding analysis placed “joy” and “anger” in the same group, producing a structure different from that of humans. The distances among emotion terms generated by humans and GPT-4 were also strongly correlated. Responses were consistent across human participants, confirming that the task captured semantic relationships among emotion terms that are broadly shared.
Importantly, these findings do not mean that GPT-4 experiences emotions in the same way humans do. Rather, GPT-4 appears to draw on relationships among emotion concepts embedded in human language to produce arrangements similar to those made by humans.
A hidden dimension emerged when 99 emotion terms were mapped
The research team next expanded the analysis to 99 emotion terms. Repeatedly asking human participants to compare all pairwise relationships among 99 terms would be impractical. GPT-4, however, can perform the same task across a large number of combinations. Both GPT-4 prompting and word embeddings divided the emotion terms primarily into two broad clusters related to pleasure–displeasure and dominance. The separation between the two clusters was clearer in the GPT-4 arrangements. A more detailed analysis also detected variation corresponding to arousal. Rather than serving as the principal axis separating all emotion terms into two groups, arousal appeared as a finer-grained dimension organizing terms within each cluster.
For example, among pleasant emotions:
“Excitement” has high arousal
“Serenity” has low arousal
A similar difference in arousal exists between unpleasant emotions such as “rage” and “sadness.” This arousal-related structure was not clearly visible when only the six basic emotions were used, but emerged after the vocabulary was expanded to 99 terms.
The number of words examined changes the emotional structure that becomes visible
These findings demonstrate that vocabulary size matters when investigating the structure of emotions. A small set of representative emotion terms can readily reveal broad distinctions, such as whether an emotion is pleasant or unpleasant, but may overlook finer differences, such as whether it is excited or calm. Some of the differences among previously proposed theories of emotion may therefore reflect the number and selection of emotion terms used in each study.
[Future Perspectives]
Using AI not as an “answer machine,” but as a measurement tool
Assistant Professor Han says, "Adapting word embeddings to a specific research purpose can require machine-learning expertise, training data, and substantial computing resources. In the present approach, by contrast, researchers give GPT-4 a task using ordinary language and analyze its responses. This method may therefore be readily accessible to researchers in fields such as psychology and linguistics who do not specialize in machine learning. Future work will examine whether the same structure appears in other languages, including Japanese, and how the arrangement of emotion terms varies across cultures."
Associate Professor Watanabe adds, "This study examined the semantic and conceptual structure of words describing emotions, rather than emotional experience itself. Further research is needed to determine how closely the structure revealed by GPT-4 corresponds to actual emotional experiences or to representations of emotion in the brain."
[Glossary]
GPT-4: A large language model developed by OpenAI. It learns from large amounts of text and generates responses to written input. In this study, GPT-4 was used to perform a task involving the spatial arrangement of semantic relationships among emotion terms.
Prompting: A method of giving an AI instructions or questions in natural language. In this study, GPT-4 was instructed to place three emotion terms on a grid according to their semantic similarity.
Word embedding: A method for representing relationships among word meanings as combinations of many numerical values. Words with similar meanings are represented as being closer together in a numerical space.
Semantic and conceptual structure: The pattern of relationships among words based on similarities and differences in meaning. This study investigated relationships among words describing emotions, rather than emotions themselves.
Basic emotions: Categories commonly treated as fundamental in emotion research. This study used six: joy, surprise, anger, fear, disgust, and sadness.
Pleasure: A dimension indicating whether an emotion is pleasant or unpleasant.
Arousal: A dimension indicating the degree of excitement or activation associated with an emotion. High-arousal examples include excitement and rage; low-arousal examples include serenity and sadness.
Dominance: A dimension indicating whether a person feels in control of a situation or, conversely, overwhelmed by it.
Scientific Reports
Mapping 99 emotion terms with GPT4 prompting reveals nuanced semantic conceptual structure.
10-Jul-2026