A new way to organize 99 emotion words — simply by prompting GPT-4
How similar are “joy” and “surprise”? And how far apart are “joy” and “sadness”? We naturally organize the words for our emotions according to how close or how different their meanings are. Yet as the number of emotion words grows, the combinations that must be compared increase sharply, making the whole picture hard for people alone to examine in detail.
Researcher Ke Han and Associate Professor Eiji Watanabe of the Laboratory of Neurophysiology at the National Institute for Basic Biology had GPT-4 arrange emotion words in space to investigate how humans organize them. As a result, “prompting” — asking GPT-4 in ordinary sentences — reproduced a structure closely resembling the arrangements made by people. Moreover, when the set was expanded to 99 words, a structure related to emotional “arousal”, hard to see with only a few words, emerged.
The study suggests that the structure of emotion concepts can be examined using nothing but questions posed in words — with no specialized AI training and no large-scale computing facilities.
Emotions can be sorted into types such as “joy”, “anger”, and “sadness”. Psychology also offers an approach that represents emotions along several continuous axes. Three of them are representative.
For example, “excitement” and “serenity” are both relatively pleasant, yet they differ greatly in arousal. “Rage” and “sadness” are both unpleasant, but again differ in arousal. Such structure has mainly been studied by asking people to rate how similar emotion words are. However, an experiment comparing all of many emotion words places a very heavy burden on participants.
The research team devised a task in which three emotion words are placed at once on a grid on screen. Words felt to be similar in meaning go close together; words felt to be different go far apart. Placing three words at once expresses three relationships simultaneously. In the first study, 89 participants and GPT-4 performed the same task using six basic emotions, and the results were compared with the conventional “word embedding” method that represents emotion words as arrays of numbers.
For the six basic emotions, humans and GPT-4 both formed two groups: “joy · surprise” and “anger · fear · disgust · sadness”. By contrast, word embeddings placed “joy” and “anger” in the same group, giving a structure different from humans. The distances produced by humans and GPT-4 were strongly correlated, and the responses among human participants themselves agreed very closely.
This does not mean, however, that GPT-4 experiences the same emotions as humans. GPT-4 is thought to have produced human-like arrangements by drawing on the relationships among emotion concepts contained in human language.
Next, the team expanded the set to 99 emotion words. Having human participants repeatedly compare all relationships among 99 words is practically very difficult, but with GPT-4 the same task can be repeated across many combinations. The analysis showed that, with both prompting and word embeddings, the emotion words split mainly into two large groups related to “pleasure” and “dominance”, and that GPT-4’s arrangement separated the two groups more clearly.
On closer inspection, GPT-4’s arrangement also revealed changes corresponding to “arousal”. This appeared not as the main axis dividing the large groups, but as an axis that finely divides the interior of each group. The structure was not clear with only six basic emotions and was detected only once the set was expanded to 99 words.
These results show that the scale of the vocabulary matters when studying the structure of emotion. Using only a few representative emotion words makes large differences such as “pleasant vs. unpleasant” easy to find, but finer differences such as “excited vs. calm” may be overlooked. Part of the reason existing emotion theories differ from one another may lie in the number and choice of emotion words used in each study.
Tuning word embeddings to a research purpose can require expertise in machine learning, training data, and a computing environment. In the present method, by contrast, researchers give GPT-4 a task in ordinary sentences and analyze its responses. This may make it an accessible method for researchers who do not specialize in machine learning — for example, in psychology or linguistics.
In the future, the team plans to examine whether the same structure appears in different languages, including Japanese, and how the arrangement of emotion words changes across cultures. Note that this study addressed not the actual experience of emotion, but the semantic and conceptual structure of the words that express emotion. How far the structure shown by GPT-4 corresponds to actual emotional experience, or to emotional representations in the brain, remains to be verified.
Ke, H., Watanabe, E. Mapping 99 emotion terms with GPT4 prompting reveals nuanced semantic conceptual structure. Scientific Reports (2026). https://doi.org/10.1038/s41598-026-60536-4