Introduction

In Western societies, blue is stereotypically associated with boys and pink with girls [1–3]. Curiously enough, these gendered associations were initially arbitrary, but became pervasive in the early 20^th century [1, 2, 4]. Nowadays, many parents continue choosing pink when dressing their daughters, decorating their rooms [5], or buying them toys [6]. Such an upbringing might explain why young girls show a particular liking for pink [7, 8]. In contrast, young adult women choose other colours as their favourite, with blue and red being the most common choices [9]. We explain these differences in colour preferences through gendered and valenced stereotypes.

Empirical studies, focused on colour preferences and colour connotations, have demonstrated that pink is considered to be a feminine colour and blue a masculine colour [3, 8, 10, 11]. Pink further represents groups of low social power and low social status [12–14]. Accordingly, adult women might shun pink to avoid being associated with these representations [9]. Red, on the contrary, represents being in power, dominant, and of high social status [15–18]. These representations potentially explain why adult women like red [9, 19–22] and why red carries both positive and negative connotations [23–27]. When it comes to valence, pink and blue both have been associated with mainly positive emotions [24, 27–29], although blue has been also associated with sadness [30–32].

When considering such gendered colours, we might first think about seeing them on girls’ and boys’ clothing, in their rooms, or their toys. Colours, however, also exist conceptually, in our minds. Colours might be gendered simply because of how we label them. If we think about pink, its gendered connotations might emerge not only because pink repeatedly occurs in feminine contexts visually, but also because it co-occurs with other feminine words in languages (e.g., pink is a girly colour). Empirical studies confirmed that colours expressed through language carry similar connotations to the same colours presented visually. Numerous studies demonstrated the red-attractiveness effect, wherein a person wearing red (man or woman) is perceived as more attractive by an opposite sex individual (for a meta-analysis, see [33]). Important here, simply mentioning that a man was wearing a red shirt had a similar effect on his attractiveness [34]. Though not focussing on the gender-loadings of colours, Jonauskaite and colleagues [35, 36] confirmed that colours, whether presented perceptually (colour patches) or semantically (colour terms), were associated with similar emotions. In the end, if the literature on embodiment and psycholinguistics holds true, an abstract meaning of a word and its respective physical representation in the world should converge on a common cognitive representation [37, 38].

We wanted to learn whether above empirical results on gendered colours can, indeed, be observed when looking at gender and valence biases in widely available written texts. Thus, we analysed gender and valence biases of colour terms in an English corpus, composed of newswire and Wikipedia articles. The latter sources both reflect and shape current standards of language use, as they are co-authored by several people and aimed at vast audiences. To extract biases, we used an artificial intelligence algorithm, focused on natural language processing (NLP). A contemporary key technique in NLP is that of word embeddings (e.g., GloVe, [39], word2vec, [40]), in which a statistical algorithm computes coordinates in a high dimensional space for each word on the basis of a reference corpus. Such NLP algorithms generally learn these coordinates from patterns of word co-occurrences in sentences. These algorithms extract not only semantic and syntactic information from everyday language, but also subtle biases in the usage of words [41]. For example, the words nurse or housekeeper frequently take a more feminine position in the semantic space than the words pilot or engineer, which take a more masculine position [41, 42].

In the current study, we used word embeddings generated from a corpus of 6 billion words by GloVe [39]. We focused on 11 basic colour terms, which included the key terms pink, blue, and red. Using these embeddings, we were able to score 100,000 most frequent words in the corpus in terms of similarities to the concepts of male, female, posemo, and negemo. The concepts of posemo and negemo, respectively, denote lists of positively and negatively laden words (see [43]). Words in our corpus that were closer to the list of words denoting each of the four concept [43] had higher similarity scores (also see, [44]). Afterwards, we computed gender and valence biases. We defined the gender bias as the difference between word similarities to the concepts of male and female, and the valence bias as the difference between word similarities to the concepts of posemo and negemo. We interpreted these biases of colour terms in comparison to i) four anchor words with clear biases (i.e., happy, sad, nun, priest); and ii) a normative word population, namely the 100,000 most frequent words in the corpus. We hypothesised that the word pink would be biased towards femininity and positivity, while the word blue would be biased towards masculinity and positivity. However, blue might be less strongly gender-biased than pink [9]. We expected red to be embedded in both positive and negative contexts, pushing its valence bias towards zero. For its gender bias, we assumed red to represent power, and informed by the literature on gender stereotypes, power would represent masculinity.

Method

We analysed gender and valence biases of 11 British English basic colour terms, namely red, orange, yellow, green, blue, purple, pink, brown, grey, white, and black [45, 46]. We also included four anchor words–two for the valence extremes (happy, sad) and two for the gender extremes (priest, nun). These words acted as sanity checks.

Results

Gender bias

We first interpreted word similarities to the concepts of male and female (see Fig 1 and Table 1). Seven colour terms red, blue, green, yellow, brown, white, and black and the anchor words priest and happy had higher similarities to the concept of male than 95% of the words in the normative word population (Fig 1A). Seven colour terms blue, green, pink, brown, white, black, purple and all anchor words had higher similarities to the concept of female than 95% of the words in the normative word population (Fig 1B). Notably, four colour terms, green, blue, brown, and black, and the anchor word happy had higher similarity scores to both concepts (i.e., male and female) than 95% of the words in the normative population.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Gender bias.

Distributions of similarity scores to the concepts of male (A) and female (B) well as gender biases (C) of colour terms, anchor words, and 100,000 most frequent words in our corpus. Colour terms and anchor words are marked with appropriate colours. Dashed lines indicate the 5^th and 95^th percentiles of the distributions (for exact values, see Table 1).

https://doi.org/10.1371/journal.pone.0251559.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Similarity scores and biases.

These are the similarity scores to the concepts of male, female, posemo, and negemo, as well as gender and valence biases of 11 colour terms and four anchor words. The position of each score in relation to the normative word population (i.e., 100,000 most frequent words) appears under “percentile”. Values in bold are below the 5th or above the 95th percentile, indicating extreme similarities or biases of these words (same data as in Figs 1 and 2).

https://doi.org/10.1371/journal.pone.0251559.t001

When interpreting gender biases (Fig 1C), the anchor word priest scored higher than 95% of the words in the normative word population, meaning it was biased towards masculinity. The anchor word nun and the colour term pink scored lower than 95% of the words in the normative word population, meaning they were both biased towards femininity. No other colour terms or anchor words had gender bias scores outside the 5^th and 95^th percentiles of the normative word population, indicating they did not have extreme gender biases (see Table 1). Thus, similarities to individual concepts (i.e., male and female) do not necessarily mean that a word has a gender bias since similarities to both concepts can be positively correlated.

Valence bias

When interpreting word similarities to the concepts of posemo and negemo (see Fig 2 and Table 1), seven colour terms, red, green, blue, pink, brown, white, and black, and two anchor words happy and sad had higher similarities to the concept of posemo than 95% of the words in the normative word population (Fig 2A). The colour term black and the anchor words happy and sad had higher similarities to the concept of negemo than 95% of the words in the normative word population (Fig 2B). Notably, black, happy, and sad had higher similarity scores to both concepts (i.e., posemo and negemo) than 95% of the words in the normative population.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Valence bias.

Distributions of similarity scores to the concepts of posemo, denoting positively laden words, (A) and negemo, denoting negatively laden words, (B) as well as valence biases (C) of colour terms, anchor words, and 100,000 most frequent words in our corpus. Colour terms and anchor words are marked with appropriate colours. Dashed lines indicate the 5^th and 95^th percentiles of the distributions (for exact values, see Table 1).

https://doi.org/10.1371/journal.pone.0251559.g002

When interpreting valence biases (Fig 2C), five colour terms, green, blue, pink, brown, and white, and the anchor word happy scored higher than 95% of the words in the normative word population, meaning they were biased towards positivity. The anchor word sad scored lower than 95% of the words in the normative word population, meaning it was biased towards negativity. No other colour terms or anchor words had valence bias scores outside the 5^th and 95^th percentiles of the normative word population, indicating they did not have extreme valence biases (see Table 1). Thus, similarities to individual concepts (i.e., posemo and negemo) do not necessarily mean that a word has a valence bias since similarities to both concepts can be positively correlated.

Discussion

We investigated whether empirical results on gender and valence biases for colour terms are reflected in an English corpus composed of newswire and Wikipedia articles. Based on previous empirical studies, we expected pink to be biased towards femininity and positivity, and blue towards masculinity and positivity. We expected red to be biased towards masculinity, while having no particular valence bias because red carries both positive and negative connotations [23–27]. To this end, we scored embeddings of the colour terms pink, blue, and red, which we obtained from GloVe [39]. Scoring was done in terms of similarities to the concepts of male, female, posemo and negemo, that means positively and negatively laden words, as defined by the LIWC project [43]. We then computed gender and valence biases as differences between similarities to the concepts of male vs. female, and posemo vs. negemo, respectively. We did the same for the remaining eight basic colour terms, four anchor words (i.e., happy, sad, nun, priest), and the 100,000 most frequent words in the corpus.

First, we checked for extreme similarities of our colour terms to the concepts of male and female. We defined extreme word similarities when a colour term had a higher similarity with a concept than 95% of the 100,000 most frequent words in the corpus (i.e., normative word population). These comparisons revealed that i) red and blue were closer to the concept of male and ii) blue and pink were closer to the concept of female than 95% of words in the normative word population. That meant that blue was close to both concepts, yielding overall no gender bias. When looking at gender biases in comparison to the normative word population, pink was the only colour term with a gender bias. We confirmed its bias towards femininity. Against our predictions, neither blue nor red were biased towards masculinity and no other colour term had a strong gender bias. Hence, we concluded that only pink conveys gender biased information (femininity) in these texts.

Second, we checked for extreme similarities of our colour terms to the concepts of posemo and negemo. Comparisons with the normative word population revealed that colour terms red, blue, and pink were closer to the concept of posemo, while no colour terms of interest were closer to the concept of negemo than 95% of the most frequent words in the normative word population. When looking at valence biases, as predicted, pink and blue were biased towards positivity and red did not have a strong valence bias. Other positively biased colour terms were green, white, and brown, while no colour term was negatively biased.

The femininity bias of pink mirrors previous empirical findings [3, 10, 11, 49, 50]. In contrast, findings on blue did not confirm a masculinity bias in our corpus, unlike previous empirical studies would have suggested [1, 3, 10, 11]. Indeed, blue seems gender-neutral, equally liked by males and females of all ages [9, 51–56]. Blue might turn into a symbol of masculinity, or boyhood, only when paired with pink, which in turn is a symbol of femininity and girliness, because pink is avoided by boys, men, and some adult women (see a more in-depth reasoning in [9, 57]). In fact, a recent study using a Stroop paradigm demonstrated that masculine words written in pink ink were perceived as being more incongruent than feminine words written in blue ink [57]. As for red, we expected, but did not find a masculinity bias, due to its associations with power and dominance [15–18]. Indeed, red is favoured by many women and might be a symbol of femininity [9, 19–22]. Thus, one can argue that red represents both masculine and feminine characteristics resulting in a negligible bias in our dataset.

Findings on valence biases were also congruent with previous empirical results. Positivity biases of pink and blue were compatible with numerous previous studies [15, 24, 27–29, 35, 36], even though blue also carries some negative connotations. Among English speakers, blue is associated with sadness in addition to several positive connotations [27, 30–32]. We did not expect, and did not observe, that red had a strong positivity or negativity bias. In empirical studies, red has been linked to positive emotions like love, joy, and pleasure as well as negative emotions like anger and hate [24, 26, 27, 35, 36, 58, 59].

Beyond our interest in (potentially) gendered colours, we confirmed only some of the previously observed valence biases for the remaining basic colour terms. Positivity biases of green, and white were compatible with previous studies [24, 25, 27, 32, 60]. However, we did not observe a positivity bias of yellow, despite numerous studies reporting associations with joy, happiness, and other positive emotions [24, 27, 29, 35, 59, 61, 62]. Unexpectedly, we observed a positivity bias of the colour term brown. In nearly all previous empirical studies, brown carried negative associations, including associations with disgust and boredom [27, 30, 35, 59]. The association with disgust is likely due to evolutionary important experiences, like rotten food and faeces, supposedly explaining why people do not like brown [52]. Experiences of rotten food and faeces might, however, be taboo subjects in widely available newswires and Wikipedia articles. Rather, people might be mentioning brown in the contexts of coffee and chocolate, both of which are positive experiences for most. Finally, we did not observe negativity biases of black and grey, despite previous studies showing that black is associated with sadness, fear, and death and grey with sadness and disappointment [24, 26, 27, 58, 60].

Our approach to detect biases in written text is methodologically different from former research studies. Maybe, these differences can explain some discrepancies between the current and previous studies. We extracted biases in an English corpus, while previous research usually tested for implicit or explicit colour meaning via individual ratings, associations, or other questioning (e.g., [3, 8, 10, 11, 15, 23–28, 60]). Additionally, when we consider the posemo and negemo lists, their names (positive emotions, negative emotions) might give rise to the idea that these words represent emotions, but most words do not describe emotions per se, but positively or negatively laden concepts (e.g., see definitions of emotions in [63–66]). For instance, sadness, anger, or joy would represent emotions, but not nice, kiss, or sceptical. Differences in definitions of concepts, study material and procedures might lead to varied study results.

Conclusions

We investigated whether pink, blue, and red have gender and valence biases in a written text corpus. With the help of artificial intelligence technology, we could show that pink was the only gendered colour, biased towards femininity. These results show that artificial intelligence can be used to assess how empirical results, often from the laboratory research studies, may relate to how people use colour terms in written texts. For some colours, our corpus study mirrored empirical study results (i.e., pink and femininity, blue and pink and positivity), for others, we observed differences (e.g., brown and positivity). Thus, we argued that written texts not only reflect human thought processes, but yield biases on their own, potentially due to selection in reporting.