Networks of emotion concepts

Abstract

The aim of this work was to study the similarity network and hierarchical clustering of Finnish emotion concepts. Native speakers of Finnish evaluated similarity between the 50 most frequently used Finnish words describing emotional experiences. We hypothesized that methods developed within network theory, such as identifying clusters and specific local network structures, can reveal structures that would be difficult to discover using traditional methods such as multidimensional scaling (MDS) and ordinary cluster analysis. The concepts divided into three main clusters, which can be described as negative, positive, and surprise. Negative and positive clusters divided further into meaningful sub-clusters, corresponding to those found in previous studies. Importantly, this method allowed the same concept to be a member in more than one cluster. Our results suggest that studying particular network structures that do not fit into a low-dimensional description can shed additional light on why subjects evaluate certain concepts as similar. To encourage the use of network methods in analyzing similarity data, we provide the analysis software for free use (http://www.becs.tkk.fi/similaritynets/).

Figure 1. The SCP method.

For detecting clusters of similar concepts, we use the SCP method, defining clusters as sets of adjacent triangles. The triangle shown in red can be rolled between five nodes in two steps (a→b→c), such that at most one node changes place at each step. These nodes constitute the cluster marked in orange in d). The three nodes marked in green form a separate cluster. The definition allows a node to belong to more than one cluster. In this example, the green and orange clusters overlap through a shared node.

Figure 2. Statistics of the similarity data.

(A) Variations observed when the same concept pair was evaluated twice. The six histograms have been constructed using each repeated pair. Red indicates the frequency P(s) of identical evaluations from most dissimilar (0) to most similar (5). Blue bars indicates the frequency of non-identical evaluations, which were typically just above or below the other evaluation. (B) Variability of averaged similarities. Each dot represents a concept pair, such that the horizontal axis indicates the mean similarity s of the pair averaged over all evaluations and the vertical axis its standard deviation. The solid curve follows an average of all points with a given similarity. (C) The similarity distribution P(s) of the Average Similarity Network (ASN). (D) The Z-score of the observed number of triangles as a function of similarity threshold, calculated using the null hypothesis of random networks with the same number of links.

Figure 3. Visualizations of the Average Similarity Networks.

Figure 4. Similarity networks of two male and two female subjects (s>0).

Figure 5. Cluster hierarchy of the Average Similarity Network.

Figure 6. Layout of emotion concepts based on a two-dimensional scaling of the similarity data averaged over all subjects.

Each concept is colored by the subjects' averaged evaluation of its (A) valence, (B) intensity, or (C) interactivity. Highlighted regions correspond to clusters determined by the SCP cluster analysis method. The smaller green regions each depict a leaf of the dendrogram. The three larger yellow regions correspond to clusters at the highest hierarchy level at which the largest cluster has been divided into a positive, negative and surprise cluster (similarity threshold s = 2.45).

Figure 7. The imbalanced triplets.

(A) The most imbalanced 16 concept triplets, in which one concept is perceived very similar to two others, whereas these two are perceived very different. The concepts are colored by their mean valence evaluations. The similarity between each concept pair is shown by link color and width, using color scales as in Fig. 3. (B) Two examples of imbalanced triplets shown on the valence-intensity (V-I) plane, where the coordinates are from mean valence and intensity evaluations of the emotion concepts. In these examples, the central concept is located between the dissimilar concepts. The concepts in panels B–D are colored by interactivity, using color scales as in Fig. 3. (C) Two more examples of imbalanced triplets on the V-I plane, where the least similar pair turns out to be located closest in the V-I-plane. (D) Concept pairs with strong similarities (s≥3.0) that are distant in the V-I-plane (Euclidian distance d≥2.0).