In Japanese culture, the light novel - a combination of text and anime-style illustrations-is a relatively new literary form. It derives from the broader otaku culture, which is also associated with video games, manga, cosplay, anime, and other forms of Japanese popular culture. Though the light novel lacks the global reach of some of these other genres, such as manga and anime, it nonetheless attracts millions of readers across a range of gender and age groups. While distinct subgenres of the light novel have emerged, such as romance, adventure, horror, and harem, issues of gender stereotyping, power imbalances and other forms of inequality remain strongly entrenched. These issues can be attributed to how otaku culture is rooted in heterosexual male desire. This paper o昀ers a quantitative assessment of these issues of gender inequality. We analyze 290 light novels, scraped from the Baka-Tsuki Translation Community Wiki, in terms of the power relationships between female and male characters as they evolve over the course of each novel. We 昀椀nd patterns consistent with issues of gender stereotyping and power di昀erentials. More speci昀椀cally, we 昀椀nd that female characters consistently wield less power than male characters, especially toward the end of each novel. We 昀椀nd some variation in speci昀椀c subgenres. We conclude with close readings of two light novels, demonstrating how a power frames approach to analyzing gender stereotypes in otaku culture augments existing work on the subject.
In Japanese culture, the light novel–a combination of text and anime-style illustrations–is a relatively new literary form. It derives from the culture of fan 昀椀ction associated with the broader otaku culture, which is also associated with video games, manga, cosplay, anime, and other forms of Japanese popular culture. Though the light novel lacks the global reach of some of these other media and cultural forms, such as manga and anime, it nonetheless attracts millions of readers across a range of gender and age groups. While distinct genres of the light novel have emerged, such as romance, adventure, horror, and harem, issues of gender stereotyping, power imbalances, and other forms of inequality remain strongly entrenched. These issues can be attributed to how otaku culture is strongly rooted in heterosexual male desire.
Otaku is a complicated term that encompasses consumer practices and fandom-related
activities as well as media and cultural production [
At the same time, the rising popularity of the light novel has resulted in a proliferation of
genres, not all of them targeted at heterosexual male readers. For instance, the shoujo (young
girls) genre is aimed at young female readers and features themes of romance and emotion
from a young heterosexual female perspective. On the other hand, the harem genre is aimed
at adult heterosexual male readers and as a result, contains more explicit sexuality. Across
these genres, a de昀椀ning feature is the light novel’s emphasis on introducing many and multiple
characters over the development of a detailed plot [
In this way, we contribute to the growing body of work that employs computational
approaches to analyze gender, including Underwood 2018, which investigates the changing
signi昀椀cance of gender in 昀椀ction written between 1780 and 2007, both in terms of authorship and
characterization [
In order to create our corpus, we scraped the English translations of 56 light novel series that were listed under the category of Light Novel (English) on the website of Baka-Tsuki Translation Community. We choose to collect our data from Baka-Tsuki because it houses the largest collection of light novels in English on the Internet, and because the translators voluntarily upload their translations to the site for sharing with the public. When scraping the novels, we employed the following criteria for inclusion: 昀椀rst, it needed to have at least one genre label; and second, it needed to have a complete English translation in a relatively standard format. We treated di昀erent volumes in the same series as separate books, in keeping with how light novels are published and read. We excluded any “side stories,” which are sometimes included in volumes of light 昀椀ction, which otherwise include stories in sequential order, because it is di昀케cult to determine whether a side story comes before or a昀琀er the main story in terms of plot, and we did not want to disrupt the main storyline. Finally, because di昀erent translators use di昀erent words to indicate a distinct volume, such as “phase” or “part,” we took these additional terms into account in the scraping process.
Our 昀椀nal corpus consisted of 289 volumes across 17 genres (16,914,599 total tokens; 98,563 unique tokens). The majority of volumes in the series were tagged with multiple genre labels, as pictured in Figure 1. While the corpus represents a small sample of all light novels ever published, most of the series included in the corpus can be considered “classic” novels, as indicated by the fact that about 71% of the series in the corpus have been adapted into anime.
A昀琀er scraping the text from the website and removing the side stories, we also manually removed all of the series that did not adhere to the volume structure, and all of the volumes that did not adhere to the one-story-per-chapter structure. (These were outliers; a typical series is broken into volumes, and a typical volume consists of 5-10 chapters). We also removed all novels in which the translator opted not to translate the text owing to its mature content. (These were indicated by the phrase, “the text may be inappropriate to people aged under 18.”) We then manually con昀椀rmed that all chapters were placed in the correct volume, and all volumes were placed in the correct order, and 昀椀xed any incorrect placement or ordering. Finally, we removed the series that contained more than 20 volumes, as we were concerned about issues of over-representation.
We used BookNLP [1] to extract the names and mentions of the characters in our corpus. We then ordered the characters by number of mentions, as we wanted to focus our analysis on the major characters of each volume. Depending on the genre of the novel, there is signi昀椀cant variation in the number of characters in each volume (with a minimum of 10 and a maximum of 193). In order to determine the cut-o昀 point between the major characters and the less important ones, we normalized the number of mentions of each character in each volume across the corpus. Below in Figure 2 we plot the sum of normalized character mentions for characters with di昀erent frequency ranks. We then picked the point of maximum curvature, which is 9, as the cut-o昀 point for the de昀椀nition of a major character.
A昀琀er identifying the nine main characters for each novel, we used BookNLP to extract the proper references and referential gender for each. Because the referential gender is inferred by the references to the characters with the text, it is sometimes inconsistent. Therefore, we used the argmax of the gender distribution provided by BookNLP to assign the gender of the character for the purposes of this project. We then further divided each volume into 昀椀ve sections of equal length (without breaking sentences), and we ran BookNLP again on each subsection-or stage-of the volume. From this, we obtained information about each of the nine main characters by matching proper references. (A match was met if the character in the subsection was mentioned by at least one of the proper references of an identi昀椀ed main character). A昀琀er locating the main characters, we extracted the actions (verbs) for which the character was agent or patient in that stage, as well as the objects they possessed, and any modi昀椀ers. We use this verb list to calculate the power score, as described in the next section.
We employed the power frames lexicon curated by Sap et al. in order to determine a power score for each of the major characters [8]. In the lexicon, each verb has a label that indicates whether the agent or the patient/theme associated with the verb has more power, or whether the two have equal power. In order to employ this lexicon, we lemmatized the verbs in the lexicon and the list of verbs we obtained for each character from BookNLP. Then, we compared the list of verbs for the characters with the lexicon, adding or subtracting from the total power score of the character as indicated. We then normalized the power scores by dividing the current score by the total number of verbs in the list. In this way, we were able to take into account the di昀erences in length and number of verbs in each novel. The 昀椀nal power score was a value between -1 and 1, with 1 indicating high power status, and -1 indicating low power status.
A notebook documenting these analyses can be found here: https://github.com/kristinagxy /lightnovel_gender
In Figures 3 and 4, we plot the mean power scores of the male and female characters in each stage of each novel, separated by genre. The bold gray line indicates the mean across the whole corpus.
As is evident in the visualizations, most of the genres follow a similar pattern, with female power decreasing over the course of the novel, and male power remaining consistent throughout. There are some outliers, however. For instance, the female mean power for adventure novels is higher than the others; and for josei (adult women’s) novels, it follows a di昀erent path altogether. This pattern seems intriguing, but upon closer look, the josei genre has the least number of volumes in the corpus. Therefore, it may not be statistically signi昀椀cant.
We were eager to evaluate the statistical signi昀椀cance of these scores. But we could not use standard methods such as the t-test, since the scores were calculated from volumes in di昀erent series, meaning that they came from di昀erent distributions. Even for volumes in the same series, the distribution was not guaranteed to be the same. To circumvent this issue, we decided to take the average of power score for each gender in each stage of each novel, and calculate a boolean score to indicate whether the average power of the female characters was larger than that of the male. In this way, we sidestep the problem of between-volume di昀erences since the resulting booleans come from a binomial distribution. Then, for each stage, we perform a one-tailed binomial test with the null hypothesis of = 0.5. The null hypothesis describes a situation of balanced power where the probability of female having a higher power score is 0.5. The p-values of the test, which represent the likelihood of seeing an outcome at least as extreme as the actual result given the null hypothesis, are reported below.
In this section, we focus on genres with interesting trends that have a reasonably large number of samples ( > 20 ). We begin with action novels ( = 155 ). As is pictured in Figure 5, at the beginning of the story, the di昀erence between the power of male and female characters is small, although the male characters’ power is always a little higher than that of the female characters. However, while the power of the female characters gradually decreases over time, the power of male increases dramatically at the end of the story, resulting in a comparatively large gap in power scores. The p-value at the 昀椀nal stage is 0.0005, indicating strong evidence that the gap exists.
In Figure 6, we show the resulting evolution of power in fantasy novels ( = 138 ). Before the 昀椀nal stage of the story, the power of male and female characters have relatively balanced values as well as high p-values, favoring the null hypothesis of equal power. However, at the 椀昀nal stage, as in the adventure genre, the two powers diverge, also with a small p-value.
The power balance in mature novels ( = 27 ) follows a di昀erent pattern. Here, in Figure 7, the power of female characters remains higher than that of male characters until the 昀椀nal section of the novel. In fact, the power di昀erential increases until the middle of each story–an intriguing trend. However the p-values for the 昀椀rst four subsections are quite high. It is only in the 昀椀nal section, when male power is higher, that the p-value indicates more than chance.
Figure 8 describes novels in the “slice of life” genre ( = 42 ), a genre typically aimed at a female readership. Interestingly, at the beginning of the story, female characters possess higher power than male characters, with a relatively low p-value of 0.100. As the story progresses, the power of female and male characters gradually become balanced, as demonstrated by high pstage stage stage values indicating that the power gap is most probably due to chance.
In the drama genre ( = 93 ), the power scores show yet another trend. Before the 昀椀nal section, the power of female characters decreases and the power of male characters increases, widening the power gap. The power gap reaches its maximum in section four, supported by a low p-value. However, by the end of the story, the two power scores converge, admittedly with a high p-value that favors the null hypothesis of balanced power.
Figure 10 shows the evolution of power in harem ( = 89 ), romance ( = 182 ), and ecchi ( = 69 ) genres. The three are very similar genres and have many overlapping titles. We observe that their trend lines are also similar, with an increasing power gap and small p-values at the end. For harem and ecchi, which are two major male-oriented genres, the power of female characters decreases consistently throughout the story.
Figure 11 shows the comparison of power across all genres. We can see that the power of female characters decreases over time, while the power of male characters remained relatively high. Interestingly, the p-value decreases as well as the story progresses, implying that the power gap is supported with increasing certainty.
To test the validity of the power frames approach to measuring power, we shu昀툀ed all the verbs for each stage and randomly reassigned them to the characters, with each character receiving stage stage the same number of verbs he/she originally had. We performed one hundred simulations for each of the 昀椀ve stages, and compared the simulated mean power scores for female and male characters across the whole corpus with the original mean scores. Speci昀椀cally, we wanted to know the likelihood that the original means came from the same distribution as the simulated means. A reasonable assumption is the simulated power scores come from a Student’s tdistribution. We obtained the p-values by calculating the probability of getting a more extreme value than the original mean using the cdf (cumulative distribution function). The results can be found in Table 7.
Interestingly, the random assignment of verbs gave a higher mean power score than the original mean for both female and male characters. For the female characters, the calculated power score in the 昀椀rst two stages does not have much di昀erence from random assignment. The later three stages clearly depart from the simulated random results, as indicated by an increasingly low p-value towards the end. This coincides with the pattern of p-value for power gap in Section 4.2. For the male characters, however, most of the time the calculated power is similar to random assignment results. Therefore, it seems that it is mainly the decreasing power of female characters that contributes to the power gap, also similar to what we observe in Section 4.2. Also notice that the simulated means for female and male characters at each stage are very close to each other, further supporting the validity of the power frames method.
What does this power gap look like in writing? To explore this question, we returned to the text of one light novel, associated with the harem genre, that exempli昀椀es the growing power di昀erential between male and female characters. The novel, the 昀椀rst volume of the Absolute Duo series, centers on Tor, a young man enrolled in Koryo Academy, a high school in which students battle each other with weapons known as Blaze. The novel opens with a depiction of the entrance test required to enter Koryo, in which Tor must battle his enemy, a young woman named Imari. He states his goal very clearly: “I will show her the di昀erence in our ability and make her give up” (Chapter 1). Here, Tor is the agent of the verbs “show” and “make,” resulting in an increase in his associated power score. This is con昀椀rmed by the power conveyed through the sentence itself, in which Tor envisions a one-way battle in which he “shows” Imari his skill and “makes her give up.” A昀琀er Tor enters Koryo, he encounters several female teachers who are more skilled than he is, contributing to the female character power score. But Tor retains his dominant position in the novel. As Tor becomes a stronger 昀椀ghter, the female characters around him grow to both trust and adore him. By the end of the novel, Tor is consistently making statements such as, “I will de昀椀nitely protect you” (Chapter 7). Again, Tor as the subject of the verb “protect” contributes to the male character power score, an indicator of the increasingly dominant role that Tor assumes in the novel.
We also wanted to explore one of the novels associated with high female power scores, even as there were few of them in our corpus. For this close reading, we turned to a josei novel, the 昀椀rst volume of the series Anohana: The Flower We Saw That Day. Anohana tells the story of six childhood friends who once formed a group called Super Peace Busters, but who grew apart a昀琀er one of the friends, a girl named Menma, died in an accident. The story is told from the perspective of a boy, Jinta, another one of the friends, who realizes that he (and only he) can see Menma’s ghost. The beginning of the novel mostly serves to introduce the characters and the backstory of the group, and contains lines such as: “That wish could only be ful昀椀lled when everyone in the Super Peace Busters was together” (chapter 3). Here, the passive construction of the verb “be ful昀椀lled” avoids assigning agency to any of the characters. As the story unfolds, Jinta tells the others in the group that he has seen Menma, but they don’t believe him. His decreasing power is represented in lines such as, “Every kind of sophisticated feeling was stuck on my body like glue, making me maximally displeased” (chapter 10). Here, Jinta is the object of the displeasure, rather than its agent. By the same token, Menma’s ghost emerges as a more powerful character in the novel, in昀氀uencing Jinta and the other friends even as they cannot see her.
In her qualitative study of gender representation in hentai (pornographic anime and manga),
Kristina Barancovaitė-Skindaravičienė (2013) argues that masculinity and femininity are
produced through bodily representation, what she describes as “indexical signs of gender” [
There is some variation based on theme, however. For genres that target female audiences (e.g. shoujo, josei, slice of life), as exempli昀椀ed by Anohana: The Flower We Saw That Day, novels o昀琀en begin by positioning the female characters with more power. As the story proceeds, however, the female characters cede some of their power. By the end, male and female characters o昀琀en have balanced power. One group of novels we did not explore were those where young girls act as heroines and protectors; most of these novels involve only female characters, and so would require a di昀erent approach to a power analysis. In the future, we hope to incorporate these as well as additional shoujo and josei novels into our study. The pattern of power in josei novels is clearly di昀erent from the majority of the other genres, but we require additional novels from the genre in order to con昀椀rm our hypothesis.