This post examines the question of whether everyday English language contains an implicit bias against men. This examination uses data reported by Dodds et al. (2011).
In a large study of reactions to the most commonly used words in the English language, words associated with females were found to evoke much more happiness and much less sadness than words associated with males. In a set of 80 commonly used gender-specific words, 75% of the 20 happiest words were female (e.g., mother, daughter) while 90% of the 20 saddest words were male (e.g., sir, man). Likewise, the female term was rated more highly — often much more highly — in 77% of matched word pairs. For instance, “women” got a score of 7.12 (where 9 was the highest possible happiness score) while “man” got only 5.94.
In their research, Dodds et al. developed a list of 10,222 words found in a massive corpus (more accurately, in four massive corpora) of English texts. The four sources from which the authors drew were as follows: 4.5 billion Twitter tweets posted over a three-year period, from 2008 to 2011; a selection from Google Books, comprising about 4% of all books ever printed; song lyrics from 1960 to 2007; and the full contents of the New York Times from 1987 to 2007.
While those sources would obviously vary in their volume and significance, it seems clear that the overall word set was comprehensive. The authors combined the 5,000 most frequently occurring words from each of those four sources to obtain their list of 10,222 words. (Of course, many of the same words would appear in multiple sources.)
The authors list those 10,222 words in a webpage, with several sets of related numbers. Although the authors do not make a clear explanation readily available, it appears that the four columns shown at the right side of that webpage are frequency rankings. For example, the word “the” is ranked No. 1 for each of the four sources: Twitter, Google Books, New York Times, and song lyrics. The four columns at the left side of that webpage present the 10,222 words and certain data related to their so-called Happiness Rank.
The authors explain that, for each of the 10,222 words, they asked 50 different people to rate how the word made them feel. The scale used for this purpose ranged from 1 (meaning “I feel extremely sad”) to 5 (“I feel no emotion”) to 9 (“I feel extremely happy”). So, again using “the” as an example, it is not surprising that its average rating, as listed on that webpage, was 4.98 — that is, very close to 5.00. “The” does not excite much emotion.
That webpage offers two other columns next to that Average rating column. One is the Standard Deviation. This is a statistical measure indicating how much variation there was among raters. It seems mildly odd that the word “the” would have much of a standard deviation. One might expect raters to be almost unanimous in giving it a 5. But apparently it speaks to some hearts.
The other column next to the Average score is the Happiness Rank. Here, again, there is a small surprise. The word “the,” with an average score of 4.98, is not exactly halfway in the list of 10,222 words. Instead, it is ranked 7,317, indicating that about 72% of the 10,222 words were ranked as being happier than “the.” The authors do not find this surprising. In another article, they suggest that, generally, human language tends to contain a bias toward positivity.
Another possibility, not explored here, is that the raters were the ones who had a positivity bias — that they were inclined and/or allowed to find more happiness than sadness in an assortment of words. If the list of 10,222 words is split in half, with 5,111 words in the top half and 5,111 words in the bottom half, the dividing line is at the average score of 5.44, rather than the 5.00 that one might expect if the raters were perfectly balanced between happiness and sadness when encountering a large set of the most commonly used words. In this light, “the” actually received a slightly negative rating, insofar as 4.98 is below 5.44 — and that seems a bit odd too.
Not all words with an average score are necessarily lacking in emotional content. Certainly many of the words achieving the median score of 5.44 (e.g., “carried”) do not seem likely to say much about happiness or sadness. But others with an average score of 5.44 could have considerable weight for numerous raters. For instance, “cope” could seem strongly positive to some, while implying an undesirable situation to others. That potential for variation is reflected in that word’s relatively high standard deviation — that is, raters varied noticeably in their reactions to “cope.”
Referring again to the authors’ disinclination to make a careful explanation readily available, the list of 10,222 terms seems to include some items that, if not outright junk, would not seem to belong there. For example, the list includes the term ‘the (i.e., with a leading apostrophe) in addition to the word the — and, for some reason, the two obtained appreciably different average scores. As a different example, the inclusion of “#p2” seems to imply that “#p2” was one of the most commonly appearing terms in at least one of the four corpora — and yet there is no ranking for that term in any of the four corpus ranking columns.
As in any study, there are certain limitations. The authors go over some of those limitations in their article. They are less inclined to discuss imperfections — but that, too, is a common feature of the research landscape. Suffice it to say that their work is sufficiently impressive to have attracted the attention of The Atlantic (LaFrance, 2016), which uses their data to analyze media coverage of Donald Trump and Hillary Clinton during the 2016 presidential election season.
Narrowing the Focus
Not surprisingly, words at the top of the happiness rank include laughter, happiness, love, and joy, while words at the bottom include terrorist, murder, cancer, and torture. And, as already noted, words like “the” and “seventy” are predictably near the middle of the pack, conveying little to no happiness or sadness to most raters.
In this post, I was particularly interested in words carrying obvious gender-related content. Hence, I assembled a list of human gender-specific terms. The entries on this list consisted primarily of synonyms of six terms (i.e., man, woman, male, female, masculine, feminine) provided by Thesaurus.com, gender-specific pronouns provided by Wikipedia (i.e., he, she, him, her, his, hers, himself, herself), words suggested by a third source, and some words (e.g., mama) that appeared in the corpora but that were not specifically provided by the sources just mentioned. From those word lists, I removed words that could be used of humans but were not human-specific (e.g., god, ape, lioness) and words that were not necessarily gender-specific (e.g., actor, which would mean a male in the context of theater, but might not be gender-specific in other contexts).
I matched the resulting list of 227 gender-specific terms with the 10,222 frequently occurring terms identified by Dodds et al. I found that 127 of those 223 terms did appear in the Dodds list. Of those 127, 75 pertained to males and 52 to females. I compared the levels of happiness that Dodds calculated for each of those 127 terms. The average happiness ratings given to the female terms was higher than the average for the males (6.35 vs. 5.95). There was a larger difference in the median happiness ratings for female and male terms (6.61 vs. 5.80). Those who have taken introductory statistics may recall that a noticeable difference between an average and a median may be due to outliers that skew the average. For example, if you have ten people, nine making $1 per year and the other making $99,991, you could say that their average income is $10,000 each — but that figure would not be very representative of the actual data. If you want to use the average (i.e., the mean), it would be more informative to discard the $99,991 outlier and say that their average income is $1. Or you could keep the outlier and use the median (roughly speaking, the fifth-highest of the ten incomes) — which, again, would be $1.
So, as I say, there appeared to be some distortion due to the presence of outliers. To reduce distortion, I removed terms that seemed likely to have strong meaning within particular contexts — to reflect, that is, on a specific situation or role rather than on an entire gender. These included in-law terms (e.g., mother-in-law, brother-in-law); religious statuses (e.g., nun, priest); political titles (e.g., emperor, queen); terms of praise or deprecation (i.e., witch, hero); and misspelled slang terms that may have suggested negative attitudes toward young delinquents (i.e., boyz, brotha, gurl). I also removed given names (e.g., Dave, Patricia), though I could instead have attempted another match-up, between the 10,222 words and a list of common first names.
Those adjustments left me with a list of 82 (45 male, 37 female) gender-specific terms. These were essentially plain-vanilla terms (e.g., man, woman), along with some of their potential variations (e.g., man’s, woman’s). And, simple though it was, this list told a story. The overall average scores for females vs. males differed markedly: 6.84 for females vs. 6.16 for males.
Among those 82 words, there was still some divergence between the median (6.84) and the average (6.68) ratings for females. I suspected this discrepancy was due to the presence of two other female-specific terms in the list. One was “ms” (more commonly spelled Ms.). That term was rated 4.48 — much lower than “mrs” (5.88). Assuming that raters were not having a sad reaction to “ms” as an abbreviation for “manuscript,” it appeared that there might be some adverse sociopolitical content to that feminist term, for at least some raters. The other term was “widow.” (Widower did not appear in the list — presumably because, in our society, women tend to be the ones who have the misfortune of having to continue to live after their spouses die.) Widow was by far the lowest-rated gender-specific term: at 2.86, in the database it was on a par with “lonely” and “mourns” and, like those terms, probably evoked sympathy rather than the rejection given to other terms receiving a similar score (e.g., illegal, violations, bombs). In other words, the ratings would be improved if the Dodds study, contrasting happy vs. sad terms, were cross-checked against a separate study of terms rated positive or negative.
Results: Significant Bias Against Males
The foregoing adjustments substantially eliminated the statistical discrepancy between averages and medians, and gave me a list of 80 gender-specific terms (45 male, 35 female) used most frequently in one or more of the four major English-language corpora listed above. As just indicated, it is not clear whether lower scores in this study would suggest sadness as distinct from negativity — but, regardless, a lower score plainly would not imply a favorable or positive attitude.
Among those 80 terms, the contrast between female and male terms was striking. For females, the median score was 6.88 (average = 6.85); for males, it was only 6.16 (average = 6.15). That gap comes out in the relative ranking of terms. Among the 20 highest-rated terms, 15 (75%) were female, whereas 18 (90%) of the bottom 20 were male. Indeed, 26 (87%) of the bottom 30 were male.
The gap also comes out in a look at pair rankings. Among those 80 terms, I identified 31 gendered word pairs (e.g., man-woman, grandfather-grandmother, niece-nephew). In these matchings of supposedly equal male and female terms, only “husband” and “wife” actually drew equal scores from the raters. Instead, for 24 (77%) out of those 31, the female score was higher than the male. Moreover, when female scores exceeded male scores, they did so, on average, by a much wider margin (0.62) than when male scores exceeded female scores (0.26). So, for example, “women” had an average rating of 7.12, while “men” had an average rating of only 5.94. Similarly, for “mother” the raters awarded an average score of 7.68, while for “father” it was only 7.06, and “girl” scored 7.00 while “boy” scored 6.24.
In a study evaluating more than a half-million rater scorings of the most commonly used words in the English language, Dodds et al. produced data demonstrating that words used to refer to human females are understood as being significantly happier than words used to refer to human males. That tendency is pronounced, and it is pervasive, applying broadly to the large majority of commonly used gender-specific terms.
These findings suggest that women rather than men are primarily associated with happiness. It goes without saying that happiness tends to be preferred over sadness, regardless of whether the precise meaning of “sad” in a specific situation is “pitiable,” “pathetic,” “regrettable,” or otherwise negative or undesirable.
The typical life consists of tens of thousands of days, each of which is typically filled with thousands of written, spoken, and mentally contemplated words. The cumulative psychosocial effect of so many words — so many of which are implicitly derogatory, for the person who does not happen to be female — can be profound.
One way to explain that impact is to use the theory of microaggressions. According to Wikipedia, that theory holds that pejorative gender-related words and acts — even if unintended or unconscious — can subtly demean the targeted individuals, portray them as aberrant or pathological, express disapproval of or discomfort with them, assume they are all the same, dispute the existence of discrimination against the targeted group, and deny the perpetrator’s bias. The concept of microaggression, described in those terms, will be familiar to many men who have been subtly targeted by the attitudes behind the pervasive linguistic discrimination detected by Dodds et al.
In an article in the prestigious American Psychologist journal, Derald Wing Sue (2007), a professor of psychology at Columbia University, explains that microaggressions need not be microassaults, which he defined as, essentially, intentional attacks. Microaggressions, he said, can also include microinsults, which “represent subtle snubs” that “convey a hidden insulting meaning,” as well as microinvalidation, which entails “communications that exclude, negate, or nullify” a person’s thoughts, feelings, or experience. While Sue’s concern was with discrimination against racial minorities, here again there are themes that will ring true, especially for disadvantaged men in today’s America, including treatment as an “alien in one’s own land,” an “assumption of criminal status” (as e.g., a potential rapist or pervert, whenever a woman or child is nearby), a “myth of meritocracy,” “pathologizing cultural values,” and “second-class status.” As Sue put it (p. 275), “The power of racial microaggressions lies in their invisibility to the perpetrator and, oftentimes, to the recipient.”
Invisible or not, such treatment can have real consequences. For instance, in her doctoral dissertation, Fay (2015) found that even a relatively small number of microaggressions — much smaller than the number implicated in the present research — was associated with increased anxiety and depression. In a similar vein, Sethi and Williams (2016) link microaggressions with adverse physical and mental health consequences. Microaggression theory is also reminiscent of laws finding a hostile work environment where discriminatory intimidation can be illegal if it is pervasive, even if it is not severe (e.g., Pospis, 2013).
Men are talked about as if they were an inferior species. This is not an occasional or isolated problem. It flows throughout the contemporary interpretation of standard words in the English language. Those who are male, or who care about males, will understand the importance of recognizing and rectifying this state of affairs. The Dodds study is not perfect. But it begins to lay the groundwork for constructive change, toward genuine equality for all people.