Investigating the structure of emotions by analyzing similarity and association of emotion words

In the field of natural language processing, some studies have attempted sentiment analysis on text by handling emotions as explanatory or response variables. One of the most popular emotion models used in this context is the wheel of emotion proposed by Plutchik. This model schematizes human emotions in a circular structure, and represents them in two or three dimensions. However, the validity of Plutchik’s wheel of emotion has not been sufficiently examined. This study investigated the validity of the wheel by creating and analyzing a semantic networks of emotion words. Through our experiments, we collected data of similarity and association of ordered pairs of emotion words, and constructed networks using these data. We then analyzed the structure of the networks through community detection, and compared it with that of the wheel of emotion. The results showed that each network’s structure was, for the most part, similar to that of the wheel of emotion, but locally different.

💡 Research Summary

This paper investigates the validity of Plutchik’s “wheel of emotion” by constructing and analyzing semantic networks of emotion words. The authors selected the 48 emotion terms defined by Plutchik (eight primary emotions and their derived secondary and tertiary forms) and collected human judgments on two distinct relational dimensions for every ordered pair of these terms: (1) similarity (“How similar is A to B?”) and (2) association (“How associated is A with B?”). Data were gathered online via a Flask‑based web application hosted on Amazon EC2, using the Japanese crowdsourcing platform CrowdWorks. A total of 360 participants initially completed each task; after rigorous quality control—including catch‑trial items and a double‑pass reliability check—27 % of similarity responses and 16 % of association responses were discarded. The final datasets comprised responses from 240 participants for similarity and 229 participants for association (average age ≈ 43 years, balanced gender distribution).

From these ratings, two fully connected weighted graphs were built: a “similarity network” and an “association network”. Nodes correspond to the 48 emotion words; edge weights are the mean rating (0–7) for the respective pair. To examine whether the networks reflect the circular organization of Plutchik’s wheel, the authors defined two quantitative indices. “Locality” measures the average normalized weight of edges linking words that belong to the same petal (i.e., the same primary emotion cluster), while “globality” measures the average normalized weight of edges linking words that sit in opposite petals across the wheel. Both indices range from 0 to 1, with 1 indicating maximal connection (weight = 7) for all relevant pairs. Results show that the association network exhibits higher locality (0.68 vs. 0.60) and higher globality (0.53 vs. 0.41) than the similarity network, and both differences are statistically significant (t‑tests, p < 0.001). This suggests that people’s sense of association spans broader emotional distances than mere similarity judgments.

For structural analysis, the authors applied Modular Decomposition of Markov Chains (MDMC), a probabilistic community‑detection method that treats a random walk on the graph and estimates, via an Expectation‑Maximization algorithm, both the community prior probabilities π(k) and the conditional node‑given‑community probabilities p(i|k). A resolution parameter α controls the granularity; setting α = 0.001 and fixing the target number of communities K = 10 yielded eight meaningful communities in both networks. The composition of these communities largely mirrors Plutchik’s petal organization but with notable deviations. For example, in the similarity network, one community groups joy, surprise, ecstasy, amazement, admiration, interest, delight, curiosity, awe, and pride; another clusters trust, anticipation, serenity, acceptance, optimism, hope, love, and aggressiveness. In the association network, similar groupings appear, yet some emotions (e.g., “annoyance” and “rage”) shift between communities, indicating subtle differences in how people perceive associative links versus similarity.

To quantify the correspondence between the detected communities and the wheel’s eight petals, the authors computed Normalized Mutual Information (NMI). The similarity network achieved NMI = 0.81 with the wheel, while the association network achieved NMI = 0.72. Both values are high, confirming that the wheel captures much of the underlying semantic structure, yet the association network is slightly less aligned, reflecting its broader relational scope.

The paper further explores inter‑community connectivity using the Ωₖ′ₖ metric proposed by Okamoto, which combines community priors, node‑given‑community probabilities, and transition probabilities derived from a PageRank‑like process. Because raw Ω values are very small, they were scaled by 10,000 for visualization. Multidimensional scaling (MDS) plots of the community‑level graphs reveal weak but systematic links, suggesting that while communities are relatively distinct, there exist meaningful cross‑community emotional pathways.

Finally, the authors examined the effect of varying the resolution parameter α, thereby extracting community structures at multiple scales. When the number of communities was forced to three, the networks collapsed into broad clusters: (i) joy‑trust, (ii) fear‑sadness, and (iii) disgust‑anger. With five communities, more nuanced groupings emerged (e.g., “grief‑pessimism‑cynicism”, “optimism‑hope‑submission”, etc.). This multiscale behavior demonstrates that Plutchik’s wheel can be interpreted at different levels of granularity, supporting its theoretical flexibility.

In summary, the study provides a rigorous, data‑driven validation of Plutchik’s emotion wheel. By contrasting similarity and association judgments, it shows that while the wheel’s overall circular topology is largely supported, local and global relational strengths differ between the two relational dimensions. The use of probabilistic community detection (MDMC) and quantitative metrics (locality, globality, NMI, Ω) offers a methodological template for future work on emotion semantics. Moreover, the findings have practical implications for affective computing and Human‑Agent Interaction: models that incorporate both similarity and association information may better capture the nuanced ways humans experience and relate emotions, leading to more natural and emotionally aware artificial agents.