Shirtless and Dangerous: Quantifying Linguistic Signals of Gender Bias in an Online Fiction Writing Community

03/29/2016 ∙ by Ethan Fast, et al. ∙ Stanford University 0

Imagine a princess asleep in a castle, waiting for her prince to slay the dragon and rescue her. Tales like the famous Sleeping Beauty clearly divide up gender roles. But what about more modern stories, borne of a generation increasingly aware of social constructs like sexism and racism? Do these stories tend to reinforce gender stereotypes, or counter them? In this paper, we present a technique that combines natural language processing with a crowdsourced lexicon of stereotypes to capture gender biases in fiction. We apply this technique across 1.8 billion words of fiction from the Wattpad online writing community, investigating gender representation in stories, how male and female characters behave and are described, and how authors' use of gender stereotypes is associated with the community's ratings. We find that male over-representation and traditional gender stereotypes (e.g., dominant men and submissive women) are common throughout nearly every genre in our corpus. However, only some of these stereotypes, like sexual or violent men, are associated with highly rated stories. Finally, despite women often being the target of negative stereotypes, female authors are equally likely to write such stereotypes as men.



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.


”Ooh! Ooh! Look at this top! I have to have it.” Nina said holding up a gold, glittery top with a boat neck but low cut back. It was gorgeous. ”And those shoes!”
Anonymous (a Wattpad story)
”I highly doubt that,” Fanny growled, clenching a hammer in her hands that she pulled from her work belt.
Anonymous (a Wattpad story)

These two excerpts from stories on an online writing community lay out competing concepts of gender roles in fiction. The first plays on common female stereotypes, while the second rejects them. Fictional worlds offer writers open-ended laboratories where they might explore new forms of social norms. Do these worlds tend to reinforce stereotypes, or push for more balanced gender roles? This paper is about quantitatively untangling signals of gender bias in an online community of amateur fiction writers.

Gender bias and its stereotypes retain significant force in modern culture. The idea of domestic and dependent women, for example, lives on in Hollywood and through subtle aspects of discrimination in science, C-level industry, and the broader workplace [Sugimoto2013, Bertrand and Hallock2001]. Similarly, the stereotype of violent and sexual men is abundant in television, video games, and popular culture [Bergstrom, Jenson, and de Castell2012, Bretthauer, Zimmerman, and Banning2007]. At the more basic level of gender representation, men are over-represented in news coverage and television shows, among other media [Ross and Carter2011, Lauzen, Dozier, and Horan2008].

Works of online fiction allow us to detect these gender stereotypes and others at scale, by examining how web users write male and female characters. The fact that such authors are amateurs helps to minimize the market incentives that drive commercial novels. What norms emerge outside these constraints? To find out, we analyze more than 1.8 billion words of fiction on Wattpad, an online community where millions of people share stories that they have written.

Four research questions guide our analysis, collectively aiming to investigate how male and female characters behave, how they are portrayed, and what reflection those characterizations have on the community:

RQ1-Action: What role does gender play in how characters act? Are male and female characters associated with different kinds of verbs?

RQ2-Description: What role does gender play in how characters are described? Are male and female characters associated with different kinds of adjectives?

RQ3-Ratings: How do patterns in character action and description impact a story’s rating? Are stories that deploy stereotypes more highly rated?

RQ4-Authors: How does author gender impact the the stereotypes we discover? Do male authors write women differently than female authors, and vice versa?

We measure gender bias by investigating stories at the level of sentences. Sentences provide a rich set of low level signals (pronoun-associated verbs and adjectives) that relate gender to character action and description. To more easily interpret these signals, we use paid crowdsourcing to group common verbs and adjectives into sixteen high level categories like “sexual” and “submissive” that capture dimensions of gender stereotypes. Over billions of words, clear patterns emerge that separate male and female characters within these categories. For example, we find that male characters are more physically active and violent, while female characters are described by terms that have submissive or childish connotations.

The size and genre variety of our corpus allow us to probe further into more nuanced questions about these stereotypes. Is there any difference in the way male and female authors write gender stereotypes? We might imagine, for example, that female authors would have a different conception of gender roles than male authors, and so write their characters differently. And what about the reception of these stereotypes? Are they associated with higher or lower rated stories? Ultimately, we find that while both genders perpetuate the same set of stereotypes on Wattpad, the relationship of these stereotypes to story rating is mixed.

Related Work

Gender Stereotypes in Modern Storytelling: Studies of gender roles in commercial media offer a useful starting point for investigating stereotypes in amateur fiction. Work analyzing movies, children’s books, and music lyrics has found that women are likely to be portrayed as younger, more emotionally motivated, less rational, and valued for beauty over intellect [Towbin et al.2008, Gooden and Gooden2001], whereas men are stronger, more violent, less in control of their sexuality, and expected to express emotion physically [Emons, Wester, and Scheepers2010, Bretthauer, Zimmerman, and Banning2007]. Women also tend to be portrayed as domestic, with activities centered around the home or family, while men are shown as physically active and economically successful, with a much wider range of employment and often higher ranking and higher paying jobs [Lauzen, Dozier, and Horan2008, Soulliere2006]. Where past studies have focused on gender bias in relatively small datasets of professionally-crafted commercial media, our work addresses the biases of amateurs across orders of magnitude more user-generated data.

Gender Stereotypes on the Web: A second body of work has found that similar gender stereotypes translate to the web. For example, males are over-represented in the reporting of web-based news articles [Jia, Lansdall-Welfare, and Cristianini2015], an effect which also holds for other forms of social media, like twitter conversations [Garcia, Weber, and Garimella2014]. Gender roles can also shape the substance of web content itself. For example, biographical articles about women on Wikipedia disproportionately discuss romantic relationships or family-related issues [Wagner et al.2015]. Sometimes gender can drive user interaction: for example, on Pinterest, women earn more repins, but fewer followers [Gilbert et al.2013]; on IMDB, reviews written by women are perceived as less useful [Otterbacher2013]; or on social media sites and dating profiles, women are more often judged on physical characteristics than men [Rose et al.2012, Otterbacher2015, Fiore et al.2008]. Similarly, in web searches for images of archetypal occupations (e.g., nurse, investment banker), the minority gender is often portrayed and perceived as relatively more unprofessional [Kay, Matuszek, and Munson2015]. Our work extends these analyses to the fictional worlds created by a community of amateur writers.

Gender and User Behavior: The gender of a user or artist can influence how they behave or what kinds of stereotypes occur in creative content. For example, on Pinterest, women more often reciprocate social links, while men specialize in curating specific categories of content [Ottoni et al.2013]. Similarly, in roleplaying games, both genders are likely to choose warriors as novices, but expert female players are more likely to default to the stereotypical healer; men, on the other hand, mainly play healers only if they design the avatar as a woman [Bergstrom, Jenson, and de Castell2012]. In television, the gender of writers can affect plot and narrative: mixed gendered groups of writers are more likely to deviate from stereotypes, writing both men and women in stereotypical feminine roles [Lauzen, Dozier, and Horan2008]. In our work, we examine how the gender of users on Wattpad affects the kinds of characters that appear in their stories.

Analysis of Text Corpora: Our approach to text analysis is informed by a large body of work in the computational social sciences. Researchers often use lexicons like LIWC to analyze texts across a broad range of signals [Pennebaker, Francis, and Booth2001]. Other researches have crowdsourced new lexical categories when they do not yet exist [Wagner et al.2015, Fast, Chen, and Bernstein2016], an approach we adopt to build categories that capture common gender stereotypes. To extract relevant signals from text, researchers have leveraged the coefficients of regression models [Gilbert2012] and the strength and significance of correlations [Pennebaker and Stone2003], as well as performed significance tests between the means of lexical categories [Kramer, Guillory, and Hancock2014]. In our work, we extend these analyses to target sentence-level relationships that capture behavioral and descriptive attributes of male and female characters.

Genre Stories Rating M. Authors F. Authors M/F Authors M. Pronouns F. Pronouns M/F Pronouns
Historical Fiction 1562 2.49 748 797 0.94 0.0098 0.0091 1.08
Vampire 6936 2.92 3423 3466 0.99 0.0111 0.0089 1.25
Teen Fiction 109873 2.84 56771 52472 1.08 0.0094 0.0084 1.12
Other 70423 3.03 38864 31109 1.25 0.0081 0.0071 1.14
Fantasy 21351 2.54 10951 10282 1.07 0.0092 0.0095 0.97
Romance 70354 2.89 36471 33455 1.09 0.0110 0.0085 1.29
Science Fiction 6193 2.63 3197 2964 1.08 0.0092 0.0070 1.31
Fanfiction 228707 2.90 125525 101789 1.23 0.0110 0.0076 1.45
Humor 11472 2.71 5982 5436 1.10 0.0065 0.0057 1.14
Mystery / Thriller 10164 2.77 5128 4966 1.03 0.0096 0.0093 1.03
Horror 7287 2.95 3815 3431 1.11 0.0091 0.0093 0.98
Adventure 11423 2.43 6088 5268 1.16 0.0087 0.0082 1.06
Paranormal 5603 2.80 2872 2696 1.07 0.0091 0.0092 0.99
Spiritual 1477 2.38 788 676 1.17 0.0073 0.0071 1.03
Action 5901 2.82 3175 2691 1.18 0.010 0.0072 1.39
Non-Fiction 8258 2.46 4390 3822 1.15 0.0053 0.0054 0.98
Short Story 11991 2.50 6353 5567 1.14 0.0087 0.0086 1.01
General Fiction 5290 2.87 2830 2431 1.16 0.0089 0.0090 0.99
Werewolf 12528 2.96 6453 6000 1.08 0.011 0.0082 1.22
ChickLit 2031 2.93 1095 923 1.19 0.0091 0.0087 1.05
All 655295 2.70 349795 301530 1.16 0.0096 0.0077 1.24
Table 1: An overview of the genres in our fiction dataset, along with high-level gender statistics. Male characters are generally over-represented across genres. The Stories header refers to the number of stories per genre. Rating refers to the average rating of stories. M. Authors and F. Authors refer to the number of stories written by men and women. M/F Authors refers to the ratio of male to female authors. M. Pronouns and F. Pronouns refer to the average number of male and female pronouns per story, normalized by story length. M/F Pronouns refers to the ratio of male to female pronouns.

Data and Methods

We take a quantitative approach to answer our research questions. Here we describe our data and statistical methods.

Data: 1.8 billion words of amateur fiction

We conducted our analyses over amateur fiction from the Wattpad writing community.111 In aggregate, our dataset contains more than 1.8 billion words selected from a random sample of 600,000 stories, written by more than 500,000 writers across twenty genres. Wattpad provided us with all data, as well as gender annotations for 475,000 authors, 46% of which are by women and 54% of which are by men. Table 1 presents a summary of the stories and genres in our dataset, as well as the gender representation in each genre, captured by counts of male and female pronouns.

How balanced are Wattpad stories between male and female characters? Before addressing our research questions, it is important to note (and control for) the fact that men are commonly over-represented in many forms of media [Ross and Carter2011, Garcia, Weber, and Garimella2014]. This representational bias also exists on Wattpad (Table 1). Consistent with prior work, male pronouns are more common across most genres: Wattpad stories use 1.24 times more male pronouns in their writing than female pronouns. Further, both genders write more about men. Male authors use 1.23 times more male than female pronouns, and female authors use 1.25 times more male pronouns (a small but significant difference, ). Our methodology adjusts for this male over-representation by normalizing our statistics between genders.

Character Statistics

Social scientists often treat documents as “bags of words,” for example analyzing tweets, books, or status updates with the lexical categories in tools like LIWC [Kramer, Guillory, and Hancock2014, Pennebaker and Stone2003]. Here we present a more targeted kind of analysis that captures the actions and descriptions of characters by taking advantage of the dependency structure of sentences.

Character behavior. Stereotypes are defined in part by how people behave. Do our stories read like Indiana Jones movies, with men initiating violence and activity, and women screaming and needing rescue? To address these kinds of questions, we mined gendered character actions from the dataset.

We extracted actions through subject-verb relationships where the subject is “he” or “she”. For example, in the sentence “she ruined him utterly” we would extract the subject-verb relationship “she ruins”. We lemmatized words in these extractions (e.g., “ruined” to “ruins”) to collapse verbs onto a canonical form.

This process left us with 42 million character actions, of which 25 million are male, and 17 million are female.

Character description. Stereotypes also exist through how people are described. Are men expected to be tall and broad-shouldered, while women are small and graceful, as cultural stereotypes would suggest? To find out, we mined gendered adjective descriptors from the dataset.

We extracted two kinds of adjective relationships. First, we targeted adjectives attached to male or female subjects by the verb “to be,” for example “he was clinically deranged.” Second, we targeted the modifiers of words that represent male or female characters like “brother” or “mother,” for example “the upset mother.” Again, we lemmatized all words.

This process left us with 4.3 million adjectives, of which 2 million are applied to female characters, and 2.3 million are applied to male characters.

Crowdsourcing a Stereotype Lexicon

Character statistics present a low level view of the differences between genders. To understand stereotypes at a higher level – for example, whether male characters more often act violently or female characters are more often described as weak – we need lexical categories that map individual words onto broader categories. Such categories would tell us that the adjective “fragile” is associated with weak, or the verb “kill” is associated with violence.

To choose these categories of stereotype, we draw from the results of prior work across television, movies, and social media [Towbin et al.2008, Lauzen, Dozier, and Horan2008]. The literature identifies the following set of categories predominantly associated with men: violent, dominant, strong, arrogant, sexual, angry, and (physically) active. For women, the predominant categories are: domestic, hysterical, childish, afraid, dependent, emotional, beautiful, and submissive. We utilize these categories in our analysis.

We built lexical categories that capture this list of stereotypes by mapping the 2000 most commonly occurring verbs and adjectives in our dataset onto the set of categories through a series of crowdsourcing tasks. Labeling the most common words in a corpus is an efficient way to build a useful lexicon while limiting cost [Mohammad and Turney2013]. For each task, we asked workers on Amazon Mechanical Turk (AMT):

For each word, tell us how strongly it relates to the topic. For example, given the topic VIOLENT, the words “kill” and“hurt” would be strongly related, the word “scratch” would be related, the word “resist” would be weakly related, and the words “laugh” and “patient” would be unrelated.

We created tasks for each category of stereotype. Each task gathers annotations for 20 words, for which we pay workers $0.14 in line with guidelines for ethical research [Salehi, Irani, and Bernstein2015]. We collect labels from three independent workers, enough to generate accurate results given high quality workers, such as the Masters workers we recruit on AMT [Sheng, Provost, and Ipeirotis2008]. To reduce the impact of false positives on our analysis, we added a word to a category of stereotype only if the majority of workers have decided it is related or strongly related to the category in question. Similarly defined tasks have preformed well when constructing lexical categories [Mohammad and Turney2013].

We generated this lexicon at a total cost of $630. Crowd workers agreed among themselves (voting unanimously for labels of related or strongly related) at a rate of 85%.

Statistical Methods

Given a lexicon that maps low level character statistics onto high level stereotypes, we now present the statistical techniques that translate these data into answers to our research questions. We used two primary methods to investigate our four research questions:

RQ1-Action and RQ2-Description

. To discover the presence of gender stereotypes in how characters act and are described, we tested for a difference in mean frequency counts between genders for our sixteen stereotype categories, taking these means over stories and normalizing them on character gender occurrence rates. We used Welch’s t-test, a two-sided test that does not assume equal variance among populations, and we present these results as odds ratios between means. Because we investigated many hypotheses at once, we applied a Bonferroni correction (

) to the results of each analysis. All the results we present reach this level of significance.

To better explain our results, and discover specific behaviors (verbs) and descriptions (adjectives) that drive our high level effects, we also tested for differences in the gendered frequency counts of 2000 individual words within our stereotype categories (e.g., how often the adjective “cocky” in the category dominant modifies male vs. female pronouns). For these word-level statistics, we again compared normalized mean frequency counts using Welch’s t-test and applied a Bonferroni correction ().

We repeated these statistical tests within individual genres to add nuance to our results and ensure that our findings were not overly influenced by one particularly large genre, like Teen Fiction.

RQ3-Ratings and RQ4-Authors

. To discover how stereotypes differ between male and female authors or across high and low rated stories, we trained logistic regression models that use gendered counts from our sixteen stereotype categories to predict either author gender or story rating. This type of model has provided insight into other kinds of linguistic signals

[Gilbert2012]. For example, if men and women write different kinds of stereotypes (e.g., if women write more hysterical

male characters, or less of them, or follow any other consistent pattern), we would expect a classifier trained on the frequency counts of stereotypes captured by our categories to have significant predictive value on author gender. Similarly, a classifier trained on these features to predict story rating will reveal what stereotypes (e.g.,

violent men or dominant women) are associated with higher and lower rated stories.

Concretely, we trained these logistic regression models using scikit-learn with L2 regularization (C=1). For both models, the input features were frequency counts of words captured by our stereotype categories for male and female characters, normalized on the baseline character occurrence rates for each gender. This made features in total. To predict author gender, we trained on two classes (male and female) over a dataset of 429,582 stories. For story rating, we split the dataset into two equal parts above and below the median rating, and trained the model to predict these classes. When reporting on classifier accuracy, we ran both models under 10-fold cross-validation. We applied a Bonferroni correction () when testing the significance of features in these models.

Finally, we computed gendered odds ratios on the 1000 most common verbs and adjectives for both male and female authors. These ratios tell us, for example, how much more likely a male character written by a man will take an action like punch, in comparison to a male character written by a women. We plotted a log-transform of these ratios for visual clarity, and computed a Pearson correlation to quantify the similarity between authors of different genders.

Figure 1: Here we present verbs and adjectives that are significantly associated () with male and female characters. The size of each word in the cloud is proportional to its odds ratio for the associated gender.


In this section we present the results of our analyses. We describe our raw statistical results here, then elaborate on them with examples in the Discussion section.

Male Odds Sample verbs and adjectives (male odds)
strong 2.02 intense (3.1), smash (2.6), intimidating (2.1)
arrogant 1.30 cocky (7.1), smirk (2.8), smug (2.6), rude (1.4)
sexual 1.22 sexiest (3.1), kiss (2.4), hot (2.1), flirt (1.5)
active 1.17 jog (2.5), lift (2.4), dodge (1.7), spin (1.4)
dominant 1.15 rich (2.8), protective (2.7), royal (2.0), command (1.4)
violent 1.10 abuse (4.4), hurt (2.3), beat (2.0), kill (1.5)
beautiful 1.06 dreamy (8.14), attractive (4.09), cute (3.3), hot (2.14)
angry 1.05 bellow (3.1), growl (2.7), curse (1.4), snarl (1.3)
Female Odds Sample verbs and adjectives (female odds)
weak 1.73 fragile (6.3), faint (3.2), sick (1.8), tired (1.4)
submissive 1.66 helpless (3.5), shy (2.9), timid (2.8), whimper (1.7)
childish 1.54 squeal (11.1), naive (7.8), giggle (4.9), silly (1.7)
afraid 1.46 shriek (4.8), frightened (2.3), shiver (1.8)
dependent 1.43 clingy (3.2), vulnerable (2.5), desperate (1.8)
hysterical 1.25 bitchy (11.4), dramatic (3.2), suicidal (3.1)
domestic 1.16 cook (2.3), wash (1.8), marry (1.7), clean (1.5)
emotional 1.04 meanest (7.9), gush (5.1), sob (3.7), fiery (2.8)
Table 2: The sixteen crowdsourced stereotype categories we used to analyze the corpus. All category odds ratios are statistically significant after correction (). Men are portrayed as more strong, arrogant, and sexual, and women as more weak, submissive, and childish.
Stereotype Direction Genres that disagree with direction
arrogant male Non-Fiction
sexual male General Fiction, Science Fiction, Historical Fiction, Adventure
angry male Fantasy, Chicklit, Teen Fiction
domestic female Spiritual
emotional female Science Fiction, Historical Fiction, Adventure, Fanfiction, Action, Spiritual
Table 3: Stereotypes are mostly consistent across genres. Here we present the few stereotype categories that show some disagreement within genres. Direction refers to the gender association of each stereotype on the whole dataset. This effect runs in the opposite direction for genres that disagree. E.g., in Chicklit women are more associated with anger than men.

RQ1 and RQ2: Behavior and Representation

All of our stereotype categories are statistically associated with either male or female characters, and these effects are significant after correction (Table 2). We find men are more often strong (2.02 odds), arrogant (1.30 odds), sexual (1.22 odds), active (1.17 odds), dominant (1.15 odds), violent (1.10 odds), beautiful (1.06 odds), and angry (1.05 odds). Women are more often weak (1.73 odds), submissive (1.66 odds), childish (1.54 odds), afraid (1.46 odds), dependent (1.43 odds), hysterical (1.25 odds), domestic (1.16 odds), and emotional (1.03 odds).

When we analyze the same categories within individual genres, we mostly find consistent agreement with the overall effects (Table 3). All genres agree on male associations for strong, active, beauty, and dominant, and female associations for weak, submissive, childish, afraid, dependent, and hysterical. For the other stereotypes, we see a handful of genres where the gender association runs contrary to the overall effect. For example, men are more associated with the emotional stereotype in Action and Science Fiction, and women are more associated with the sexual stereotype in Adventure and Historical Fiction.

Investigating individual words within our stereotype categories, we find 409 significant associations, of which 222 are associated with female characters, and 187 with male characters (Figure 1). While the significant words for male characters are more or less evenly divided between verbs (53%) and adjectives (47%), the significant words for female characters consist mostly of adjectives (72%).

Positive with rating Coef. Negative with rating Coef.
sexual (male) 2.03 strong (female) -0.96
arrogant (male) 1.45 domestic (male) -0.66
sexual (female) 1.24 afraid (female) -0.66
violence (male) 0.92 weak (male) -0.63
active (male) 0.90 domestic (female) -0.57
hysterical (male) 0.59 strong (male) -0.51
hysterical (female) 0.57 dominant (female) -0.44
anger (male) 0.56 emotional (female) -0.44
violence (female) 0.46 beautiful (female) -0.39
childish (male) 0.42 weak (female) -0.34
angry (female) 0.20 dependent (female) -0.25
emotional (male) 0.12 childish (female) -0.21
submissive (female) 0.02 active (female) -0.03
Table 4: Thirty out of thirty-two categories of stereotype are significantly associated () with positive and negative story ratings in a logistic regression. Sexual and arrogant men are the strongest predictors of a high rated story. Strong women and domestic men are the strongest predictors of a low rated story.
Figure 2: Male and female authors write very similar characters, with 0.98 Pearson correlation () over the odds ratios of verbs and adjectives ascribed to male and female characters. In blue we plot the odds ratios that hold the same direction across male and female authors (94%), and in orange the ones that shift direction (6%). The lower left quadrant consists of verbs and adjectives both men and women ascribe to female characters (e.g. dainty, squeal). The upper right quadrant consists of verbs and adjectives that both men and women ascribe to male characters (e.g. cocky, thrust).

RQ3: How do stereotypes affect story ratings?

A logistic regression model that we trained on normalized stereotype counts for each character gender predicts story rating at an accuracy of 88%. As this model makes predictions over two classes (stories above and below the median rating), chance accuracy is 50%. In other words, gender stereotypes are strongly associated with story ratings. We present the categories of stereotype that are statistically significant for this model in Figure 4.

Positive associations with rating. Sexual, violent, hysterical, and angry stereotypes are positively associated with story rating for both male and female characters. Arrogant, active, childish, and emotional stereotypes are also positively associated, but only for male characters. Submissiveness is positively associated with rating, but only for female characters.

Negative associations with rating. Domestic, strong, and weak stereotypes are negatively associated with story rating for both male and female characters. Afraid, dominant, emotional, beautiful, dependent, and active stereotypes are also negatively associated, but only for female characters.

RQ4: Do men and women write genders differently?

A logistic regression model that we trained on normalized stereotype counts for each gender predicts author gender at an accuracy of 53% (chance accuracy is 50%). If male and female authors wrote different kinds of gender stereotypes in their stories, this model would perform much higher than chance. So, it appears that both men and women write indistinguishably stereotypical genders. While this model provides little practical predictive value, one feature of the model did achieve significance after correction: the active stereotype for female characters is weakly associated with male authors. This means that men are slightly more likely to write female characters who take physical actions (verbs like “chase” or “lift”) in their stories.

In our second analysis, we find a Pearson correlation of 0.98 over odds ratios that describe the likelihoods of character actions and adjective descriptors written by male and female authors. 94% of these odds ratios agree in their direction (Figure 2). For example, both male and female authors associate punching with male characters, with 2.13 and 2.0 odds respectively. Alternatively, a small number (6%) of verbs and adjectives shift direction between male and female authors. For example, male authors are more likely to apply the adjective “immortal” to female characters, whereas female authors are more likely to apply it to male characters. The data subsets for male and female authors uniformly support the high level relationships in Table 2, with slight deviations in the magnitudes of these trends. For example, female authors are slightly less likely to apply hysterical adjectives to female characters than male authors, but overall still quite likely to do so.


We asked whether amateur writers, limited only by their imaginations, might contradict conventional gender stereotypes. To this question, it seems the answer is largely no. We also investigated whether (forward-thinking) genres or author genders are more likely to buck the trend. The answer, again, seems to be no.

RQ1-Behavior: How do characters act?

The statistically strongest male-oriented actions confirm gender stereotypes like arrogant, active, violent, angry, and sexual. Male characters more often smirk (2.7 odds) and mock (1.6 odds), dodge (1.5 odds) and chase (1.6 odds), abuse (4.4) and kill (1.5), roar (4.3 odds) and curse (1.6 odds), or kiss (2.4 odds) and lick (2.0 odds). For example, “he narrowed his eyes and mocked her laughter,” “he dodged attack after attack,” “he mentally cursed,” or “he tried to lick her ear.” Similarly, the statistically strongest female-oriented actions reinforce stereotypes through categories like childish, hysterical, submissive, afraid or domestic. Female characters more often giggle (6.7 odds) and bounce (2.7 odds), squeal (10.3 odds) and stomp (2.0 odds), cook (2.3 odds) and wash (1.8 odds), cling (4.0 odds) and whimper (2.3 odds), or shiver (3.2 odds) and scream (2.2 odds). For example, “she bounced up and down in excitement,” “she would help clean the compound with all other women in the clan,” or “she clung to him for a while.”

While none of our high-level stereotype categories run directly against behavioral stereotypes, we do observe a surprisingly weak male association with angry (1.05 odds). Despite the “angry man” stereotype, many common behavioral terms in this category are associated with female characters, for example, stomp (1.7 odds), huff (1.4 odds), force (1.3 odds), and storm (1.3 odds). In contrast, the male associated behaviors in this category tend to have a darker connotation, like beat (2.0 odds), rip (1.7 odds), or snarl (1.3 odds).

RQ2-Description: How are characters described?

The most statistically male-oriented adjectives confirm gender stereotypes through associations like strong, arrogant, and dominant. Male characters are more often described as intense (3.1 odds) and intimidating (2.1 odds), cocky (7.1 odds) and rude (1.4 odds), or rich (2.8 odds) and famous (2.5 odds). For example, “this guy was so intimidating,” “he was a typical cocky Californian guy,” or “his rich and famous dad.” Likewise, the most statistically female-oriented adjectives reinforce stereotypes like submissive, dependent, hysterical. Female characters are more often described as helpless (2.4 odds) and fragile (5.9 odds), clingy (4.4 odds) and vulnerable (2.1 odds), or bitchy (12.2 odds) and dramatic (2.1 odds). For example, “she went weak in the knees,” or “like all overly dramatic teenage girls, Vanny got over her breakup,” or “she was too clingy, so he dumped her.”

Against traditional gender stereotypes, we find that beautiful is significantly male associated. This effect size is weak, however, and both genders are often physically objectified in our corpus. Women are more often described as beautiful (3.3 odds), flawless (1.6 odds), curvy (46.0 odds), or slim (3.5 odds). For example, “her pale skin was nearly flawless.” But men are described with their own comparable set of terms, including gorgeous (1.3 odds), hot (2.2 odds), dreamy (3.0 odds), and toned (19.3 odds). For example, “she watched him walk, shirtless and dangerously gorgeous.”

RQ1 and RQ2: Stereotypes among genres

Gender stereotypes are mostly consistent across genres, suggesting they are not driven by any one particularly popular sub-community on Wattpad. For example, when we first discovered the male association with beauty, we wondered whether slash fiction (a popular sub-genre of Fanfiction focused on gay relationships) might have biased our results. The genre analysis provides strong evidence to the contrary, as the beauty association held across every genre in our corpus. Similarly, we find that ten of the fifteen stereotype categories we analyze are completely consistent across genres.

Beyond providing evidence that stereotypes generalize across Wattpad, the genre analysis opens up a series of new questions. For example, why are women more associated with sexual stereotypes in General Fiction, Science Fiction, Historical Fiction, and Adventure stories? Does this deviation give female characters more or less agency than we see in other genres? Similarly, why are men more emotional among Action, Adventure, and Science Fiction stories? On inspection, it appears that strong emotions emerge in the course and aftermath of physical conflict, disproportionately undertaken by male characters in these genres. Better characterizing these sub-communities remains future work.

RQ3: Story rating and stereotypes

While gendered category counts strongly predict story rating, conventional stereotypes are not always associated with higher rated stories. In support of conventional stereotypes, sexual, arrogant, violent, angry, and active men and hysterical, submissive women are positively associated with story rating. Similarly, domestic, weak men and strong, dominant women are negatively associated with rating. However, these results are tempered by contrasting trends. Sexual, violent, and angry women are also associated with higher rated stories (albeit with weaker coefficients than their male counterparts), as are childish and emotional men.

Four stereotype categories are significant for only one gender. Afraid, dependent, and beautiful female characters are associated with lower rated stories, and arrogant male characters are associated with higher rated stories. Similarly, two stereotype categories flip between genders. Active and childish male characters are associated with higher rated stories, whereas active and childish female characters are associated with lower rated stories.

Taken as a whole, these results present a mixed account of the how gender stereotypes are received on Wattpad. Clearly, they exist in the community. But they may not always be encouraged by readers. Many common stereotypes, like sexual and violent men, have a positive impact on rating, but others, like beautiful women, have a negative impact.

RQ4-Authors: Author gender and stereotypes

Authors of both genders are responsible for the stereotypes we see in the Wattpad dataset. Our first analysis suggests that patterns of gender stereotypes in a story cannot be used to predict author gender. And our second analysis shows that men and women write extremely similar kinds of characters. Of the 2000 verbs and adjectives we analyzed for association with male and female characters, few of them disagree between author genders (the orange dots in Figure 2).

Of these points of disagreement between male and female authors, 40% involve words in our stereotype categories. A handful of these words present instances in which women reject gender stereotypes. For example, female authors are more likely to apply the adjective “strongest” to female characters, and “weary” to male characters, while male authors do the opposite. Others of these words speak to a sense of difference between genders that lies outside conventional stereotypes. One group of the statistics is self-deprecating within gender. For example, men are more likely to describe male characters as “flustered,” “embarrassed,” or “chubby,” while women are more likely to apply the same adjectives to female characters. Another group of the statistics aggrandizes the opposite gender. For example, women are more likely to describe male characters as “smarter,” “magical,” or “sparkling,” while men more often apply these adjectives to female characters.


Our approach has limitations, which in turn provide opportunities for future work.

Sentence level analysis. In this paper, we analyze language at the level of sentences, examining verbs and adjectives for bias based on gender. While the concreteness of this approach allows us to efficiently draw out quantitative insight, it also excludes the greater context of the story. An independent female character dying after leading a rebellion, for example, tells us something different about gender bias than the death of a submissive and beautiful female character. Looking beyond the sentence level to story structure could allow us to consider how stereotypes interact with each other, or relate to earlier plot events in a story.

External Validity. Our work asks whether modern writers, free from market incentives, would reinforce or contradict gender stereotypes in their stories. There are several ways in which our results may not generalize to fully answer this question. First, while the Wattpad writing community is worth studying in its own right, it may not accurately represent the broader community of writers, amateur or otherwise. Second, writers on Wattpad may not be entirely free of commercial incentives: some writers have used the community as a launching point for their careers in commercial fiction, and many writers desire as wide an audience as possible. However, given the immense scale and diversity of the Wattpad dataset (more than 1.8 billion words, split almost evenly across male and female authors), we suggest that it provides the best available means towards answering our questions.


We ask whether modern amateur writers, limited only by their imaginations, reinforce or contradict gender stereotypes. To find out, we analyze more than 1.8 billion words of fiction on Wattpad, an online writing community. We find that modern amateur writers do reinforce traditional stereotypes, and that male and female authors are just as likely to perpetuate them.


First, we would like to thank Jordan Christensen and Mohammad Islam at Wattpad – this work wouldn’t be possible without their help and collaboration. We would also like to thank our reviewers and colleagues at Stanford for their helpful feedback. This work is supported by a NSF Graduate Fellowship.


  • [Bergstrom, Jenson, and de Castell2012] Bergstrom, K.; Jenson, J.; and de Castell, S. 2012. What’s ’choice’ got to do with it?: Avatar selection differences between novice and expert players of world of warcraft and rift.
  • [Bertrand and Hallock2001] Bertrand, M., and Hallock, K. F. 2001. The gender gap in top corporate jobs. Industrial & Labor Relations Review 55(1):3–21.
  • [Bretthauer, Zimmerman, and Banning2007] Bretthauer, B.; Zimmerman, T. S.; and Banning, J. H. 2007. A feminist analysis of popular music: Power over, objectification of, and violence against women. Journal of Feminist Family Therapy 18(4):29–51.
  • [Emons, Wester, and Scheepers2010] Emons, P.; Wester, F.; and Scheepers, P. 2010. He works outside the home; she drinks coffee and does the dishes: Gender roles in fiction programs on dutch television.
  • [Fast, Chen, and Bernstein2016] Fast, E.; Chen, B.; and Bernstein, M. 2016. Empath: Understanding topic signals in large-scale text. In Proc. CHI 2016.
  • [Fiore et al.2008] Fiore, A. T.; Taylor, L. S.; Mendelsohn, G.; and Hearst, M. 2008. Assessing attractiveness in online dating profiles. In Proc. CHI 2008.
  • [Garcia, Weber, and Garimella2014] Garcia, D.; Weber, I.; and Garimella, V. 2014. Gender asymmetries in reality and fiction: The bechdel test of social media. ICWSM 2014.
  • [Gilbert et al.2013] Gilbert, E.; Bakhshi, S.; Chang, S.; and Terveen, L. 2013. ”i need to try this”?: A statistical overview of pinterest. In Proc. CHI 2013.
  • [Gilbert2012] Gilbert, E. 2012. Phrases that signal workplace hierarchy. In Proc. CSCW 2012.
  • [Gooden and Gooden2001] Gooden, A. M., and Gooden, M. A. 2001. Gender representation in notable children’s picture books: 1995–1999. Sex Roles 45(1-2):89–101.
  • [Jia, Lansdall-Welfare, and Cristianini2015] Jia, S.; Lansdall-Welfare, T.; and Cristianini, N. 2015. Measuring gender bias in news images. In Proc. WWW 2015.
  • [Kay, Matuszek, and Munson2015] Kay, M.; Matuszek, C.; and Munson, S. A. 2015. Unequal representation and gender stereotypes in image search results for occupations. In Proc. CHI 2015.
  • [Kramer, Guillory, and Hancock2014] Kramer, A. D. I.; Guillory, J. E.; and Hancock, J. T. 2014. Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences 111(24):8788–8790.
  • [Lauzen, Dozier, and Horan2008] Lauzen, M.; Dozier, D.; and Horan, N. 2008. Constructing gender stereotypes through social roles in prime-time television.
  • [Mohammad and Turney2013] Mohammad, S. M., and Turney, P. D. 2013. Crowdsourcing a word-emotion association lexicon. Computational Intelligence 29(3):436–465.
  • [Otterbacher2013] Otterbacher, J. 2013. Gender, writing and ranking in review forums: a case study of the imdb. Knowledge and information systems 35(3):645–664.
  • [Otterbacher2015] Otterbacher, J. 2015. Crowdsourcing stereotypes: Linguistic bias in metadata generated via gwap. In Proc. CHI 2015.
  • [Ottoni et al.2013] Ottoni, R.; Pesce, J. P.; Las Casas, D. B.; Franciscani Jr, G.; Meira Jr, W.; Kumaraguru, P.; and Almeida, V. 2013. Ladies first: Analyzing gender roles and behaviors in pinterest. In Proc. ICWSM.
  • [Pennebaker and Stone2003] Pennebaker, J. W., and Stone, L. D. 2003. Words of wisdom: language use over the life span. Journal of personality and social psychology 85(2):291.
  • [Pennebaker, Francis, and Booth2001] Pennebaker, J. W.; Francis, M. E.; and Booth, R. J. 2001. Linguistic inquiry and word count: Liwc 2001. Mahway: Lawrence Erlbaum Associates 71 2001.
  • [Rose et al.2012] Rose, J.; Mackey-Kallis, S.; Shyles, L.; Barry, K.; Biagini, D.; Hart, C.; and Jack, L. 2012. Face it: The impact of gender on social media images. Communication Quarterly 2012.
  • [Ross and Carter2011] Ross, K., and Carter, C. 2011. Women and news: A long and winding road. Media, Culture & Society 33(8):1148–1165.
  • [Salehi, Irani, and Bernstein2015] Salehi, N.; Irani, L. C.; and Bernstein, M. S. 2015. We are dynamo: Overcoming stalling and friction in collective action for crowd workers. Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems 1621–1630.
  • [Sheng, Provost, and Ipeirotis2008] Sheng, V. S.; Provost, F.; and Ipeirotis, P. G. 2008. Get another label? improving data quality and data mining using multiple, noisy labelers. Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining 614–622.
  • [Soulliere2006] Soulliere, D. 2006. Wrestling with masculinity: Messages about manhood in the wwe.
  • [Sugimoto2013] Sugimoto, C. R. 2013. Global gender disparities in science. Nature 504(7479):211–213.
  • [Towbin et al.2008] Towbin, M. A.; Haddock, S. A.; Zimmerman, T. S.; Lund, L. K.; and Tanner, L. R. 2008. Images of gender, race, age, and sexual orientation in disney feature-length animatedfilms. Journal of Feminist Family Therapy 2008.
  • [Wagner et al.2015] Wagner, C.; Garcia, D.; Jadidi, M.; and Strohmaier, M. 2015. It’s a man’s wikipedia? assessing gender inequality in an online encyclopedia. In Proc. ICWSM 2015.