In the history of feminism and gender theories, there have been different attempts to categorize gender. From a stereotypical binary categorization of people as either male or female ([de Beauvoir 1992], [Bourdieu 2010]) up to a negation of biological indications of gender [Butler 2003], there have been many thoughts on what it could be that defines humans as belonging to one or another category of gender. A pioneering contribution that is still relevant today (cf. [Connell 2015, 61, 64–5]; [Kirkpatrick 2020]), is The Second Sex by Simone de Beauvoir. In this article, we analyze this approach because it represents and reproduces a large set of stereotypical female roles (e.g. “wife,” or “working woman”) by differentiating them from male counterparts. We use this binary, clearly defined dichotomy as a starting point for defining two opposite categories and questioning this opposition. We analyse this text in particular because it is still one of the most relevant and often cited texts with regard to gender aspects. Additionally, it marks a turning point in the study of gender: Although Beauvoir analyzes stereotypical roles, she shows some roles that do not fit into a binary understanding of gender. More importantly, she points out the social construction of all of these roles and the features most typically assigned to them. Especially the famous sentence “one is not born but made to be a woman” is often quoted as a key statement. However, a systematic analysis of all gender roles contained in the text and the transfer into a model that can be used for further analysis of gender roles does not exist at this point. In her study, Beauvoir extracts gender roles from literary texts, among others, to illustrate what kinds of gender roles exist. To transmit the gender roles presented and features assigned to them by Beauvoir into a transferable category system we annotated the second part of the text.[1] We first develop a gender model based on Beauvoir’s theoretical text. Beauvoir herself quotes indirectly from literary texts or uses direct quotations in order to show which gender roles exist. In a second step, we apply the gender model highlighted by Beauvoir to the literary texts she cites. Since gender roles alone are not the only important factor when it comes to gender aspects, we also look for features that are attributed to different gender roles in the texts. We do this to further differentiate the model and ultimately gain a more accurate understanding of how gender roles are constructed in literary texts. Our contribution can be understood as a transfer of a non-digital study to computational literary studies. As Beauvoir uses a common approach of non-digital humanities we wanted to redesign the attempt to a controlled setup in which exactly the same amount of text is used from each of the sources. Instead of listing some of the quoted features we gather all of the roles and features assigned to characters in these passages. We then compare the gendered characterizations in a network analysis. Moreover, we assemble all characters’ roles, clothes and features in one graph in order to see whether there are common features assigned to certain gender roles.

Beauvoir ends up with an intrinsically binary system of female and male roles[2] and gives examples of a few alternative gender roles (such as “lesbian”). Social acceptance of certain features plays a key role for her attempt. She describes how certain features[3] (e.g. narcissism) are either supported (for female children in the case of narcissism) or oppressed (for male children in the case of narcissism) during youth depending on the gender identity that shall be formed. It is important to keep in mind that the focus here lies on the development of certain features according to social norms, whereas individuals can be born with any kind of feature combination. It is also crucial to note that Beauvoir does not end up with a fixed set of features for each role she describes. Instead gender roles turn out to be defined as fuzzy sets (cf. [Zadeh 1965]) because

1. gender roles are not clearly distinguishable categories but show permeable borders and sometimes even overlap, and
2. gender roles are not defined by a fixed set of features that all individuals embodying such a role show. Instead individuals can show a subset of features assigned to a gender role and even some features assigned to another role.

In addition gender roles are non-exclusive, which means that a character can show more than one gender role.

What holds true for the gender roles developed by Beauvoir is also true for the gender categories we come up with through analyzing her work: in each gender category various gender roles are included; some of them are more closely related to a top-level-category than others; roles that are not very closely related to the top-level categories can sometimes also be grouped into another top-level category. We do not exactly know which features form a category, where a category ends and a new one begins. We do not assume a closed, rigid theoretical gender model, but rather a network of commonalities that can overlap and intersect, thus yielding diverse and less sharply separated gender roles.

We mainly use two methods from the spectrum of Digital Humanities: digital manual annotation and network analysis. With annotation, we chose a method that is central to hermeneutic text analysis and which is commonly used in the (Digital) Humanities (cf. [Jacke 2018]). However, annotation is most often used in the analysis of primary sources. In our approach, we annotate literary texts, as well as Beauvoir’s The Second Sex. In order to make a list of all literary works she cited in the second part of her work and to connect all features named by Beauvoir to the gender roles she assigns them to, we annotated the second part of her work, in which she elaborates on her system of gendered roles. In a close reading process we thus annotated (a) literary works she cites and (b) features and typical clothes she describes for the different gender roles. We then did a context-analysis of the feature annotations in order to relate features and clothes to gender roles. Altogether we used an expert annotation approach in which one of us first did the whole annotation and then in a second runthrough the other one conducted the context-analysis and corrected annotations if necessary. In this second runthrough works, features, clothes and gendered roles were matched as the data was manually transferred to the network analysis software Gephi ([Bastian et al. 2009]). In a next step we turned the annotation of cited works into a list and built up a corpus containing some of these sources.

Having annotated Beauvoir’s text we turned to these primary sources and annotated features, clothes and roles used to describe characters in a sample of 19 novels (see section 2.2). We used the same close reading expert annotation approach as described above and built one database for each of the texts. To further analyze character features in the novels we used OpenRefine ([Krause 2021]) to disambiguate the features so that shared features of characters show more easily (see section 4). We then imported the data to Gephi to visualize it in one single networked graph.

Being a multidimensional network a graph can show different kinds of entities and relations. We ran a force-directed network algorithm that visualizes nodes with many relations to each other or with many shared relations to other nodes using proximity (cf. [Schumacher 2018]). Nodes that are neither connected to each other via direct nor indirect relations are shown as being located very far from one another. The algorithm was stopped when equilibrium was nearly reached, i.e. as the nodes and edges stopped changing their location. We thus compare the creation and functioning of a category system in two different research settings.

In the following part, we describe the category system that emerged from the conceptual annotations of Beauvoir’s theoretical opus magnum. In her text, Beauvoir describes gender roles by referring to sociological observations and by using examples from historical, psychoanalytic, and literary texts ([Botond 2017, 117]). In doing so, she undermines the division between disciplines and employs an approach that Sigmund Freud, for example, uses as well:[7] real-world oriented, person-related observations, phenomena, and roles substantiated through characters in numerous yet exemplary literary texts. Thus she develops a category system of gendered roles such as “mother,” “father,” “young girl,” “young man,” and so on. Her argumentation always follows descriptions of the female roles first. In most cases, male roles are developed accordingly. Comprehensibly, the focus on the female physique and psychology leads to a far more detailed description of female roles than of male roles. Further on, Beauvoir gives examples of roles that cross the border between male and female e.g. “virile woman,” “tomboy,” “feminine man” or “dandy.” As these roles are not only characterized by traditionally female traits but also by male characteristics they incorporate diversity. As can be seen from Figure 1 Beauvoir indeed describes the roles “woman” and “man” as opposing each other. But some other, more unexpected particularities show when interpreting the graph data:

1. Although “woman” and “man” clearly oppose each other they do not show the largest distance in the graph. “Girl” is situated behind “woman” and “son” appears behind “man.”[8]
2. By contrast, roles incorporating diversity are situated in the middle of the graph, but
   • linguistically female roles[9] (“lesbian,” “virile woman,” “tomboy”) are situated towards the male pole of the graph, whereas
   • linguistically male roles[10] are situated closer to the feminine pole (“feminine man,” “gay man,” “dandy”).
3. There are some cases of transcendence of feminine roles towards the male pole and vice versa:
   • “independent woman” is situated in close proximity to “man,”
   • “Wife” is situated in close proximity to “husband,”
   • “The crucified” is situated even behind the role “woman.”
4. Neutral roles are underrepresented in the graph and are not situated exactly in the middle.

Altogether the graph representing gender roles as described by Beauvoir in The Second Sex appears to be quite spherical: male and female areas are transferring each other, especially female roles cover a wide stretched area inside the graph. As expected, female roles are more dominant in number than male roles are.

In this chapter we will describe the postprocessing of the annotations in our fiction corpus and shed some light on the distribution of features and roles within it. We mainly focus on collective and individual features and roles but this time they are bound to individual literary characters or groups of characters instead of the gender roles they might stand for.[13] To be able to interpret the large amount of annotation data, we disambiguated all annotations. We performed the removal of linguistic ambiguities semi-automatically as well as manually using OpenRefine.[14] Disambiguation involves different forms of unification at a grammatical and a semantic level.[15]

After disambiguation, a quantitative evaluation of the annotations shows that a total of 4,558 features,1,283 roles and 519 clothes were annotated in the entire corpus. Among these features, two broad areas can be identified: collective features and singular features. We define collective features as features that occur with high frequency (i.e., repeatedly and more than ten times, meaning that more than ten characters are described by these features). By singular features, we mean features that occur only once in the entire corpus and therefore are also used to describe one single character only. Both groups can be further categorized into internal and physical features. Internal features are not visible from the outside and describe the personality of a character. Physical features describe the appearance and the physiognomy of a character and are visible from the outside.

In this part of the paper we bring together the annotation data from the fiction corpus (i.e. the features, clothes and roles of the characters) with the gender categories developed using the data gathered from Beauvoir’s The Second Sex. In order to handle the amount of data and keep the setting comparable we use Gephi ([Bastian et al. 2009]) as network visualization software again and stick to the same algorithm used for the gender sphere of Beauvoir’s The Second Sex, named Force Atlas 2. It is not possible in the scope of this paper to fully discuss roles, features and clothes and their respective use to build character profiles. Instead we use all three parameters to model the graph and to base positions of characters inside it on all three dimensions. Having already analyzed collective and singular features from the annotation data, in the analysis of the graph we highlight gender roles in order to proceed with the categorization of gender. From previous research (cf. [Flüh et al. 2022a], #fluh_etal2023, [Flüh et al. 2022b] we know each character in a literary text is usually associated with more than one single gender role. Protagonists are often characterized by a large number of different gender roles. So in addition to features and clothing we annotated all depictions of characters as potential gender roles. Features, clothes and roles were assigned to individual characters thus building an individual character profile. This assignment is based on close readings of the text chunks features, clothes and roles appear in. We classified the roles according to the basic gender categories “female,” “neutral” and “male” in a first step as we did in the first part of this study. The color code is the same we used in Figure 1: green for “female,” blue for “neutral” and red for “male.” Features, clothes and roles are the most important basis for the positions of characters in this network. In addition to that work titles hold together characters of one text a little. In order to focus on gender categories we gave all but roles nodes a gray color here which turns them almost invisible. Focussing on the gender roles three things become apparent:

1. In contrast to the data from the non-fiction corpus in the fiction corpus male gender roles are more dominant than female roles,
2. nevertheless female and male roles are more balanced in this corpus than in the non-fiction corpus,
3. in the center of the graph representing the fictional corpus a zone of neutral roles is situated.

We conclude that in contrast to the graph data of the non-fiction corpus the graph of the fiction corpus shows a scale. However we need to look more closely in order to see if and where diversity finds its place in this corpus. But we would like to stress the fact that neutral roles are way more important in the fiction than in the non-fiction corpus. This meets our general impression that neutrality is often overlooked in gender theories.[18] From the graph data representing the fiction corpus (cf. Figure 12) one can see that even in historical data one can frequently find roles associated with more than two genders, mostly being either female or male and neutral. But it is less common to find profiles that show an equal amount of or more or even exclusively neutral roles than such that include either male or female roles only.[19]

The fact that we are dealing with characters described using features, clothes and roles in the fiction corpus and not with roles being described by features leads to a different organisation of the graph: roles are situated less towards the margin and more towards the center. It is due to this characteristic that the most frequent male, female and neutral roles show in the middle of the area they are situated in. Zooming into the graph one can see that in the fiction corpus transcending roles, i.e. roles being classified as one gender category in an area mostly occupied by roles of another category, are appearing very frequently. Nevertheless, mostly these transcending roles are represented by very small nodes as they are only used to describe a single character. As we did before we exclude these cases from our categorization as we do not think this to be enough data to base a valid interpretation on. Once again the category “neutral” provides us with some special cases which we will analyze in some detail later on. But first we turn towards the category “male.”

By zooming in on “man” as the most frequently used male role to describe characters in the fictional corpus, we can identify roles most closely connected to the center of the male area of the graph as well as characters mostly described using this or similar roles and works in which these highly frequent male roles are used especially often (cf. Figure 13).

Highly frequently used roles that can be found in close proximity to “man” are e.g. “brother,” “father,” “(male) friend,” “mister” and “lad.” Characters that show many such roles in their profile are e.g. Julian from Red and Black and Fjodor Petrowitsch Karamasow from The Brothers Karamasow. Novels in which a high number of these roles appear are – unsurprisingly – Red and Black and – more strikingly – The Portrait of Dorian Gray. By looking at graphs we modelled using data of single texts only we can validate that indeed most of the characters introduced in the first 20,000 tokens of these two novels show mostly male roles in their character profiles. Actually in both novels there is only one character that shows a high number of features as well as female gender roles in their profile as can be seen from Figure 14.

By gradually zooming out of the graph we sorted the most frequent gender roles into the top level category “male.”

What can already be seen from Figure 13 is that “human” as an intrinsically neutral category shows up in close proximity to the role “man.” This means that in fact the role is more often used to describe characters that show a high number of male roles in their character profile than for roles that show a higher number of roles of other genders. We will see that there are equivalents of this phenomenon in the female area of the graph later. But first we would like to zoom in on the center of the graph where neutral roles like “child” can be found (cf. Figure 15).

Again by gradually zooming out we identified roles for the top-level category “potentially neutral” from the preliminary category system we developed earlier. Roles that are situated here are “person,” “being,” “darling,” “parents,” “intruder” (being the neutral “Eindringling” in german), “guardian angel” and “audience.” Roles in this category could be further grouped into terms used for single persons and groups, as nicknames or swear words. Characters that can be found in close proximity of these roles are Sibyl Vane from Dorian Gray or Madame Volmar from Juliette. Novels that are situated close by are War and Peace as well as Juliette. In contrast to the findings from the analysis of the male area this does not mean that there are especially many characters in these novels that show neutral roles in their profiles. As Figure 16 shows Juliette indeed features a main character that is attributed with many neutral roles whereas War and Peace has a comparatively more balanced number of male and female roles used to describe main characters.

The most central and highly frequent female roles in addition to “woman” are “Lady,” “girl,” “daughter,” “(female) friend,” “widow,” “nun,” “mother” and “sister” (cf. Figure 17). Characters that can be found closely to the center of the female area of the graph are Atala and Juliette from the novels of the same names. There are no novels that are positioned so closely next to the role “woman” as Dorian Gray and Red and Black are to “man,” which we see as an indicator that there is no novel constructing some kind of “women’s world” in the opening passage such as Dorian Gray takes its beginning in a clearly male-dominated setup. But there are Pot Luck by Balzac and Well of Loneliness by Hall as two novels that can be found not far from the center of the female area of the graph.

The graphs representing the data of these two novels show that Pot Luck indeed has more characters that are mostly characterized by female roles than characters mostly described using male roles. The Well of Loneliness is a very special case because it contains more characters with mostly female roles in their profile as well as one character that shows female, male and neutral roles in an almost balanced way. This is the main character Stephen who we found to be one of the very rare characters in this corpus that can be called “diverse” in terms of gender (cf. Figure 18).

Altogether for male as well as neutral and female roles, family relations are an important factor. Similar to the male area yet less apparent is the transcendence of a frequently used neutral role into the female area of the graph. Whereas characters mostly described by male roles are highly frequently also referred to as “human,” characters that are described using mostly female roles are often also depicted as “creature” (although the German word “Geschöpf” actually found in our data is much less negatively connotated). Another gender reference transcending from the neutral area into the female sphere through frequent use in mainly female character profiles is “thing.” So although all three terms – “human,” “creature” and “thing” – are intrinsically not bound to a certain gender the frequency of use in combination with male and female roles differs significantly. However it can be seen from the whole graph that many neutral minor roles, i.e. less frequently used to describe characters, can be found all over the graph. Due to the focus of this study which does not lie on the analysis of gender neutrality we have to postpone further studies of the subject to future research.

Finally we turn to the question whether there are roles used inside this corpus that could be classified as rather female or rather male. If we use the preliminary category system developed in part one we come to a disappointing insight. Most of the roles suggested by Beauvoir are not used in this corpus to describe characters. No “gay man,” no “lesbian,” no “independent woman.” Only two roles of the category “rather female” can be identified being “dandy” and “painter.” This is not a very surprising finding, when we keep in mind that we are dealing with historical data from the timespan 1353–1945. But does this mean that not any gender diversity is present in this corpus? Bringing together our finding that Stephen from Well of Loneliness is profiled using a balanced amount of female, male and neutral roles with the fact that Beauvoir mentions the heroine of this novel in the section of her book called “the lesbian” (cf. [de Beauvoir 1992, 855]) we assume that gender diversity (except for the very rare cases in which “dandys” are named as such) might not be explicitly mentioned in our fiction corpus but profiled using a variety of roles of the other gender categories. Once again, we are facing the problem of fuzziness. Whereas we can see from the graph data of the fiction corpus that it is very common that characters show a majority of either male or female roles together with some neutral roles and one role of the opposite gender (when assuming a gender scale as we do from analyzing the graph data), roles that show at least two roles of the opposing gender in their profile are more rare. Finally characters like Stephen that show a profile with an almost balanced amount of roles of three gender categories are exceptions. While we are certain that such a profile can be used to describe a character that does not fit into the boundaries of a binary notion of gender without stating it explicitly, the question whether it would suffice for a character to show two gender roles of the opposing gender (in relation to the majority of gender roles in its profile) in the first 20.000 tokens of a narrative must be postponed to further research in which we analyze them as whole texts.[20] However there is one character in our corpus showing a very similar profile as Stephen, which is Jo from Alcott’s Little Women (cf. Figure 19).

Jo is not named as “lesbian,” “virile woman” or “independant woman” by Beauvoir but rather described as a tomboyish young girl, who nevertheless gets married in a traditional constellation. Yet Jo shows a strikingly similar profile as Stephen. Both characters are linguistically constructed as female, i.e. referred to with she/her pronouns. Among others both characters are described with the central and highly frequent male role “man.” Stephen is also referred to as “boy” and Jo as “brother.” Another interesting similarity is that both characters are referred to by using male historical figures, being Nelson and Wilhelm Tell in the case of Stephen and Shakespeare for Jo. Both characters are also described using names of male animals such as “cock” (Stephen) and “stallion” and “colt” (Jo). Although Beauvoir does not name Jo as an example for a character of a non-binary gender category we take from this that we are dealing with a character that clearly breaks the boundaries of such a notion of gender. This is especially remarkable as the novel dates from 1868 and thus shows that gender-profiles of characters that clearly do not fit into the boundaries of a binary notion of gender can be found as early as in 19th century fiction.

Coming back to the analysis of the whole graph we can now answer the final question of where to find these characters that contest binary notions of gender in the graph. Other than in the graph representing the non-fiction corpus, characters that show potential for gender diversity are not situated between the (neutral) center of the graph and the male and female areas towards the poles. Instead they show up beside the main axis of the scale situated towards the margins of the graph (cf. Figure 20).

Although this is not nearly enough data to come to a final conclusion about the categorization of gender diversity from here we can use the position inside the graph as a heuristic hint towards characters that are of potential use for further analysis of the topic. Such characters are for example Dorian Gray from the novel of the same name and Felix de Vendenesse from the Lily of the Valley. In order to proceed with the categorization of gender diversity one could annotate and analyze these novels as whole texts and analyze the resulting gender profiles.

The revisited category system looks very different from the preliminary one we developed in part one of this study (cf. Figure 21 to Figure 8).

We kept the categories “female,” “neutral” and “male” in which we could sort many of the roles named and categorized by Beauvoir. However we did not find equivalents to the roles she depicted for “rather female” and “rather male.” Actually only two of these roles were named in the fictional corpus being “dandy” and “painter.” We realized that characters that do not fit into a binary gender system are not explicitly named with roles of “rather male” or “rather female” gender. Instead they show high numbers of roles of all three gender categories “female,” “male” and “neutral” in their profile. We thus came up with another category which we named “diverse.” As the former category “potentially neutral” it is very different from the other categories as it shows only a low number of specific roles. Again it is only a first attempt to gender diversity that can be used to further proceed along this path. It suggests that diversity should be analyzed at character level using an individuation-based approach rather than on the level of roles. As can be seen from Figure 21 neither the non-fiction nor the fiction corpus provided us with enough data to categorize all gender roles. By using the same colors as in Figure 8 we visualized categories we kept from the first attempt, those we moved to another category and those we added (colored in white but framed in the color of the top level category). The visualization also shows that once again many roles could not be classified as they were not named in the fiction corpus. Finally the visualization neither includes all the low frequency roles from the fiction corpus as it would simply be too much to include more than 1,000 roles in one visualization.

By annotating theoretical as well as literary texts in terms of features, clothes and roles used for characterization we were able to reconsider Beauvoir’s approach to gender roles and to implement it into a category system. We mapped the gender sphere she creates in her theory and were able to show that gender roles incorporating diversity are well integrated into the center of this sphere. In addition we could single out that neutral roles are underrepresented and almost entirely overlooked in her approach. However her notion of gender varies considerably from a scalar notion of gender ranging from male to female and having a neutral center. Such a notion can be modelled using the data of the fiction corpus. In this notion of gender diversity is set at the margins, i.e. appears as falling outside of the scale.

By analyzing distinctive/individual and overarching/collective features we realized that Beauvoir highlights features that according to her are gender-specific but that are not very common in the fiction corpus. We also found that highly frequent features are mostly used in a gender-unspecific way in the fiction corpus whereas features with a lower degree are more likely to be part of gender-specific patterns. However we conclude that gender roles are of higher importance for profiling gender in the fiction corpus. By analyzing individual gender profiles we found out that characters not fitting into a binary gender system are characterized using an almost balanced amount of female, neutral and male roles. But we also found that it is very common for literary characters to show a majority of roles of one gender plus an indefinite number of neutral roles and a single role of the opposing gender (when following the notion of a gender scale). However our data was not sufficient to show the exact breaking point from characters that fit into gender norms to those that break them – although we have been able to single out characters and works that could be of interest for further research on this subject.

In the end we came up with two category systems that can be used to further investigate gender characteristics in literary texts. One closely follows Beauvoir whereas the other is based on gender roles, features and clothes frequently used in a corpus of fiction. By transferring Beauvoir’s non-digital approach to a controlled computational research setup we could show the differences in modelling gender in one way or the other. After having tested Beauvoir’s gender model on the primary sources she used herself and adjusted the resulting category system, we can now transfer this categorization of gender to other research areas such as contemporary literature.