The Association of Color and Emotion
By Marlena Tyldesley
In 2019, a Swiss team of data analysts and computer scientists set out to better understand the relationships we experience between our emotions and the colors around us. Our world, they point out in their introduction, is an unbelievably colorful place. Not only that, but every day we choose to add more color to it; clothing, makeup, wall decor… we add color to our lives anywhere we can. The colors we choose are “rarely random,” the authors say. They can be attributed to many different things, including the emotions we associate with them. This color-emotion association is what the authors hoped to better understand in their research.
The authors of this study developed a machine-learning approach to understanding the link between colors and emotions, and whether those associations are similar across cultures. To do so, 711 participants from 4 countries (China, Germany, Greece, and the UK) completed an online survey about their own color-emotion associations. The survey employed the Geneva Emotion Wheel, pictured here:
The Geneva Emotion Wheel is a tool used to assess the degree to which each participant associated a given color (RED, in this example), with 20 different emotions. The outermost circles represent the strongest emotional association with the color, while the innermost square represents no emotional association with the color. The participants were also offered the option of selecting “no emotion,” meaning that the color carries no emotional association whatsoever for them, or “different emotion,” where participants could add their own emotion. “Different emotion” submissions were not considered for this study.
Over the course of the survey, participants completed a wheel for each of the 12 color terms assessed. It is worth noting that the participants were shown the word for each color in their native language (hence, “color term”), not a picture of the color in question. Participants also provided demographic data including country of origin, native language, and fluency in the language they took the survey in (only participants fluent in the survey language were included).
Now comes the heart of this study, the machine-learning employed to analyze the survey data. The authors used statistical learning (also known as multivariate pattern classification) to predict the color a given participant would have associated with each of the 20 emotions. The basic process is this:
10 participants are randomly chosen from the group
The survey responses from 9 out of those 10 participants are fed into the statistical classifier (the authors’ algorithm). The data from that group of 9 is referred to as the training set.
At this point, the statistical classifier has information from 9 people about how they associate color with emotion.
The 10th person’s data is the test set. For each of the 20 emotions in the test set, the authors used the classifier to predict what color the 10th person would associate with them.
So, the statistical classifier has “learned” the general emotion-color preference of the group it was initially given. For example, let’s say that in the first group of 9 people, 8 of them strongly associated anger with RED. Now, when we ask the classifier what color it predicts the 10th person will associate with anger, it would likely tell us RED. On the other hand, if 3 out of 9 associate anger with RED, 3 associate it with BLACK, and 3 associate it with BROWN, the classifier will probably be less accurate in predicting the 10th person’s color. By that logic, the more accurate the classifier is in predicting that 10th person’s response, the more consistent that color-emotion association is in the population.
This is a highly versatile approach to data analysis, because that group of 10 can have anything in common that the authors choose. For example, they could choose 10 people from China and see how consistent answers are among Chinese participants. Or they could choose a much larger sample, and see if responses are consistent among, say, all 711 participants.
The authors conducted this experiment in two sections. The first was designed to address the hypothesis “If emotion associations are highly colour-specific and consistent across participants, it should be possible to predict at high accuracy which colour terms a participant had evaluated, given the 20 emotion ratings.” The second concerned differences between color-emotion associations across cultures, and addressed the hypothesis “If emotions between colour and emotion are country-specific, then it should be possible to predict a participant’s country of origin using their set of 240 colour-emotion association ratings… accuracy of the statistical classifier is proportional to the country-specificity of the colour-emotion associations.”
In the first section, the authors found that the strongest association across the board was between the emotion love and the color RED. Furthermore, the statistical classifier’s accuracy (and therefore the consistency of participants’ answers) was highest for BLACK and RED, followed by BROWN, PINK, and GREY, all of which “elicited relatively specific emotion associations.” ORANGE was frequently classified incorrectly as YELLOW, which suggests that participants associated them with similar emotions.
In the second section, the authors found that, in fact, color-emotion association was fairly similar between participants from different countries. There were, however, some differences to note. BROWN, for example, was most strongly associated with disgust in Germany. This association was essentially not present in China. PURPLE was only associated with sadness in Greece - possibly because darker shades of purple are sometimes worn in mourning there - while WHITE was only associated with negative emotions in China, possibly because that is the color traditionally worn to funerals in that country. YELLOW was generally associated with positive emotions in all of the countries except Greece.
While these results are limited and cannot reliably be applied to entire countries, they are a good indication of the power of machine-learning in research across disciplines. In fact, the authors note that, while their use of machine learning is novel in this field of study, it is well-known and frequently used in neuroscience. Not only does this study offer insight into the topic which it studied - emotion and color associations in various cultures - but also into the potential of machine-learning as a research tool in the future.
Jonauskaite, D., Wicker, J., Mohr, C., Dael, N., Havelka, J., Papadatou-Pastou, M., Zhang, M., & Oberfeld, D. (2019). A machine learning approach to quantify the specificity of colour-emotion associations and their cultural differences. Royal Society Open Science, 6(9). https://royalsocietypublishing.org/doi/10.1098/rsos.190741#d3e2429