Compared to the French, Italian, and Spanish theatre traditions, which had normalised female actors in the sixteenth and early seventeenth century, the Dutch were relatively late in allowing women on the stage. Ariana van den Bergh (1626/1628-1661) famously broke the taboo in 1655, when she played her namesake Ariane in Onvergelijkelijke Ariane (1644), a play by Jan Jacobsz Schipper. She became a celebrity immediately after her first performance in the Schouwburg, paving the way for other Dutch actresses such as Susanna van Lee, Elisabeth de Baer, Alida Molensteen, Emerentia Veltzen, Joanna Vissers, Anna de Prendre, Maria Besems, Cornelia de Vlieger, Van den Bergh’s daughter Maria Nozeman, and Van Lee’s daughter Adriana Eeckhout.⁸ Many actresses had marital or family ties to professional male actors.⁹ Women had occasionally performed on Dutch stages before 1655, usually as singers or in travelling theatre companies, but Van den Bergh’s debut in the spotlights of the prestigious theatre at the Keizersgracht marked a clear break in Dutch theatre history.

However, Ariana’s stage debut did not mean that female characters were no longer embodied by men after 1655. The stage continued to be dominated by male bodies, either in male or female dress. Reading the Personageboek, a unique source which documented the distribution of roles among actors in the theatre season of 1658-1659, Louis Grijp concluded that in this year the main female roles were usually played by the professional actresses Van den Bergh, Van Lee, and Kalbergen. Other female characters, however, were mostly played by men. Genre seemed to have played a role, since male impersonators of female characters were more common in farces than in tragedies, possibly because of the more sexual nature of those roles.¹⁰ Furthermore, cross-dressing was not limited to male actors: in some cases, Van Lee and Van den Bergh played minor male roles.¹¹ There is also evidence that female roles could be played by either a male or a female actor depending on the staging, which again confirms the flexible approach to gender roles during the early years of Dutch female acting. Even in singing roles, there was not a necessary connection between the gender of the characters and the gender of the actors. In the seven stagings of Vondel’s Gysbreght van Aemstel (1637) in the 1658-1659 season for example, the dramatic Rey van Klaerissen, voiced by a group of nuns whose lament of the biblical Massacre of the Innocents allude to their own cruel rape and murder reported in the fifth act, was sung by male singers.¹²

The introduction of female actors not only affected the theatre’s stage, but also its ticket office. In the archives of the Schouwburg in Amsterdam, digitised in the ONSTAGE project, Olga van Marion and Frans Blom found evidence of a steep increase in the theatre’s annual income after Ariana van den Bergh’s debut on stage in 1655.¹³ The opportunity to see women performing apparently drew big crowds to the Keizersgracht. By counting the female roles in several plays, Van Marion concludes that their share increased after female actors entered the stage, possibly as a result of the financial incentive to create more space for women in the story worlds of plays.¹⁴ She also tested (but could not confirm) the hypothesis that female playwrights were more inclined to attribute more lines to female characters than their male colleagues.¹⁵ In this article, we will complement her findings with new statistics on the visibility and characteristics of female voices on stage.

Besides female actors, female authors also played a role in the changing performances of female identities. A few female playwrights, including Katharyne Lescailje, Catharina Questiers, and Catharina Verwers Dusart, claimed their position in the male-dominated domain of the theatre by writing plays that were staged in the Schouwburg. In her readings of the first Dutch plays authored by women, Amanda Pipkin recognises a tendency among female playwrights to rewrite the moral codes concerning sexual misbehaviour. Through their dramatic and poetic criticism of gender relationships, authors such as Verwers and Questiers rehabilitated women and rejected ‘the most injurious stereotype that women were the source of all sexual immorality’.¹⁶ Martine van Elk has shown how Verwers and Questiers in their plays negotiated early modern ideals about public and private femininity, in dialogue with the sources of their adaptations, which originated in the work of Jacob Cats and Spanish authors Antonio Enríquez Gómez, Félix Lope de Vega, and Miguel de Cervantes.¹⁷ Johanna Ferket contributed to this debate by analysing the female roles in (male-authored) Dutch farces, arguing that their female characters often engage in different discourses than those of their male counterparts. Even in farces, female characters were not limited to sexual stereotypes but often mimicked female roles and female positions in everyday life and everyday discussions, for example allowing ‘women to speak out on stage about their roles as mothers’.¹⁸

These readings highlight the way both female and male playwrights used the stage as a space where gender relationships and early modern ideals of femininity and girlhood could be questioned and reimagined.¹⁹ Meanwhile, the increasing presence and participation of women before and behind the scenes seems to have affected the role of Dutch female characters and the representation of gender relationships. But how exactly? We lack an empirical understanding of the historical consequences of this increased visibility and participation of women in the theatre. Does a more equal (or less unequal) gender balance lead to more pronounced distinctions between gender roles, or, conversely, to more fluid boundaries between male and female identities? After the gradual normalisation of female actors on the Dutch stage in 1655, gender performativity possibly became more rather than less fluid. While female characters were always men in female dress before the 1650s – which may have guaranteed a certain stability and conventionality in the performance of gender – in the second half of the seventeenth century male and female bodies could be impersonated by men or women. There must have been stagings where female characters played by men interacted with female characters played by women, alongside male characters embodied by men, and female actors in male dress. The common dramatical use of disguises and false identities thus acquired a new layer of complexity, especially in the many comedies and other plays that made gender switching a driving force of their plot. A well-documented example is the role of Rozaura in the Dutch adaptation of Pedro Calderón de la Barca’s La vida es sueño (1636), titled Sigismundus, prinçe van Poolen (1654). In later stagings of that play, Ariana van den Bergh first played a male traveller in drag, then switched to a court lady, and finally switched again to a (male) soldier.²⁰

This flexibility in the performance of gender roles leads us to formulate two, contradicting hypotheses: it made gender roles on the Dutch stage either more stereotypical (to compensate for their lack of clarity), or less pronounced (because apparently gender was not so strictly defined). As a historical and comparative question, this research problem requires a large sample that is illustrative, if not representative, of dramatic productions in the studied period. Our selection of ninety-eight plays is too limited for that purpose, but it offers a solid starting point for further analysis from both a historical and a transnational perspective. We consider our analysis of the relationship between gender and character speech in early modern Dutch drama as a first step towards a more comprehensive history of changing gender roles on both Dutch and European stages.

The corpus was extracted from DutchDraCor, a collection of tei-encoded editions of early modern Dutch drama that contained ninety-eight plays at the moment of extraction (April 2024).²¹ All selected plays were first printed between 1613 and 1786. DutchDraCor is one of the subcorpora integrated into the Drama Corpora Project (DraCor), an infrastructure and ‘programmable corpus’ that facilitates (cross-lingual and longitudinal) computational analysis of theatre corpora in several European languages.²² Each play included in DraCor has been encoded in all elements from the hierarchical dramatic structure: acts, scenes, speeches, and lines. All speeches are marked by a character id and all characters are labelled with a gender code: female, male, unknown, or other.

Our selection includes the largest sample of fully encoded Dutch plays currently available, even though it contains only a fraction of the total production of Dutch drama from the seventeenth and eighteenth centuries. There is no complete overview of all Dutch plays printed in this period, but there must have been several hundred. Mieke Smits-Veldt has identified more than 180 discrete tragic plays published in Amsterdam or by Amsterdammers between 1600 and 1650.²³ Her estimation was based on Hubert Meeus’s Repertorium van het ernstige drama in de Nederlanden 1600-1650, which documented no fewer than three hundred plays from the Low Countries in this period.²⁴ The online database of the Amsterdam City Theatre from the aforementioned ONSTAGE project records 999 distinct Dutch plays first printed in the period studied in this article (1613-1786).²⁵ We consider a sample of ninety-eight plays (ca. 10 percent of the relevant plays included in the ONSTAGE database) sufficient to develop an informed hypothesis about the relationship between gender and speech in Dutch drama in the seventeenth and eighteenth century. In the future, our analysis will have to be replicated and complemented from larger samples.

The corpus contains plays from six different genres: sixty-three tragedies, seventeen comedies, eight farces, five pastoral plays, three tragicomedies, and two morality plays. All genre labels were derived from the titlepages or other paratexts from the original editions. Eighteen plays (18.4 percent) were written by female authors: seven by Katharyne Lescailje, seven by Lucretia Wilhelmina van Merken, three by Catharina Questiers, and one by Catharina Verwers. The plays are not equally distributed over the studied period: the current state of DutchDraCor has a bias towards the seventeenth century (fig. 1).

We extracted all speeches from all plays and grouped them by character and by gender, excluding characters with gender code ‘unknown’ or ‘other’. This resulted in a dataset of 37,976 speeches by 1141 characters or single-gender character groups (such as reien, choruses). A large majority (70.3 percent) of the characters is male, which also means that male speeches outnumber female speeches by a factor of 2.2 (tab. 1).

In theory, character speech can be gendered in various ways: male characters might use different words than female characters; they could act in different social situations and therefore speak about different topics and in different discourses; or they could speak in different styles. Although these textual features are likely to overlap to a certain degree – a different style usually leads to different lexical preferences – we differentiate between the lexical, semantic, and stylistic level of character speech in order to capture the complexity of potential gender differences in dramatic language. Each level requires its own computational approach, focusing on relative frequencies of content words (vocabulary), semantic connotations of words and sentences (semantics), and patterns of the most frequent (function) words (style).

In both the lexical and semantic approach, we approach the question whether there is a relationship between the gender of a character and their speech as a classification task. We use machine-learning techniques to predict or determine the gender of a character in a text based solely on the language used within that text. This is a popular method in the field of quantitative text analysis and has been applied to several contexts. Examples are gender identification of the authors of blog posts, tweets, e-mails, and forum messages.²⁶ In the context of literary studies, the study of Hota et al. on gendered speech in Shakespeare theatre texts is noteworthy.²⁷ More recent studies in which a similar task is executed include the work of Sven Vitse, who analysed gendered character descriptions in a corpus of modern Dutch novels.²⁸

The central question is thus whether we can predict the gender of a character based on their speech text. Our first attempt to answer that question focused on lexical preferences by characters, or simply: their vocabulary. We trained and evaluated a Term Frequency-Inverse Document Frequency (tf-idf) classification model using logistic regression. With tf-idf, the importance of a word is determined by weighing its occurrence in the document and computing how often the same word occurs in other documents. A word frequency matrix is built of all words in the corpus, in this case of all the words in all speech texts. Prior to that, some preprocessing steps are taken to prepare the data.

The procedure is as follows. First, we group the speeches of every character. Next, we make chunks of equal size from these speech texts, experimenting with different chunk sizes. This means that if the speeches of character X have a total length of ten thousand words, they are chunked in twenty fragments of five hundred words, or in twenty-five fragments of four hundred words. With this step we prevent bias being introduced to the model by the varying lengths of speech texts. Punctuation is removed from the speech texts, after which the fragments are lemmatised (converting all inflected words to their lemma or dictionary form) to decrease the number of unique words in the corpus and thus reduce the size of the tf-idf matrix.

A balanced dataset is created, meaning there are as many speech texts by male speakers as female speakers. The size of the dataset depends on the chunk size – larger chunks result in a smaller dataset. We create random samples for both genders, using 80 percent for training and 20 percent for testing. The TfIdfVectorizer from the Python package scikit-learn is used to transform the text data, and a LogisticRegression model is trained on the training set across fifty iterations. The trained model is then used to make predictions on the test set.

The semantic approach builds on the lexical approach, with the important difference that this time texts are represented using a transformer model, also known as a bert model. This is a type of machine-learning model that processes text and generates representations, called embeddings, which capture the meaning and structure of documents. These embeddings encode both the semantic content – what the text is about – and stylistic features, such as tone or writing style, allowing the model to understand and compare documents based on deeper linguistic patterns. This makes transformer models highly effective for tasks like clustering or categorising texts. There are many different transformer models, each trained on specific types of data and optimised for various tasks. The variation in training data means these models can capture different nuances, ranging from everyday language to technical or historical styles, depending on the specific context they were designed for.

To decide which model suits our task and corpus best, we perform a clustering task in which we test five potentially suitable models, including one historical Dutch language model, two contemporary Dutch language models, and two multilingual models.²⁹ The clustering task includes the following steps: we restrict our corpus to the forty characters with the most speech text, and chunk their speech text into fragments of similar length. We experiment with chunk sizes of two hundred, three hundred, and four hundred tokens (word instances). We create embeddings and apply K-means clustering to see with which embeddings the text fragments of each character cluster together the best. The performance is measured using the V-score, a clustering evaluation metric that measures the harmonic mean of homogeneity and completeness. We work from the assumption that the model with the highest performance is the most suitable to encode the semantic and stylistic features of early modern Dutch theatre text. The results in tab. 2 show that the m-e5-large model consistently performs either best or second best across all experiments. However, when the chunk size is reduced to two hundred tokens the GysBERT-v2 model outperforms the m-e5-large model. In order to compare the performance of both models, we continue our analysis with m-e5-large and GysBERT-v2.

Preparing data for fine-tuning a bert model requires pre-processing steps that differ slightly from the pipeline described in the lexical approach. Because bert takes context into account, we do not remove punctuation, nor do we perform lemmatisation. The texts are tokenised and encoded using the pre-trained bert model to generate contextual embeddings. These embeddings are then used as input features for a Logistic Regression model, which is trained on the training set across fifty iterations. As in the lexical approach, the trained classifier is used to make predictions on the test set, enabling a direct comparison of model performance across representations.

Besides the lexical and semantic dimensions of character speech, we tested whether a character’s gender was marked by the stylistic features of their lines. In the field of literary studies, there are myriad definitions of what literary scholars tend to call ‘style’. We specifically follow the conceptualisation used in the field of computational stylometry, which views literary style not only in qualitative terms, but also from a formal and quantitative perspective.³⁰ Computational literary scholars have developed and tested several ways to operationalise their focus on quantifiable features of style, but in general they measure style as patterns of word frequencies in a given text that are characteristic for an author or character. They continue to engage in vibrant arguments about the distance metrics and parameters that best fit a given question, as they tend to agree that there is no one-size-fits-all approach to stylometric problems in general.³¹ We decided to use one of the oldest and most generic distance metrics, Burrows’s Delta, because it often functions as a benchmark and its general applicability across different genres and use cases has been demonstrated sufficiently.³² Burrows’s Delta was designed to normalise the weight of individual words regardless of their frequency, which enables a balanced stylometric analysis of the vocabulary despite the unequal distribution of words in any text. Stylometric distance was computed based on a small selection of the vocabulary: the hundred most frequent words (mfw) of the combined vocabulary of all plays selected for analysis. We decided to use a relatively small section of the mfw list compared to other applications of Burrows’s Delta, because plays are relatively short texts compared to novels, with an average length of 12,273 tokens (character speech only), and because they are segmented in (even shorter) selections of male and female character speech.³³ We used the R package ‘Stylo’ developed by Eder et al. to compute the stylometric distance between all male character speech and all female character speech in a given play from a given oeuvre by four playwrights: Jan Harmensz. Krul, Joost van den Vondel, Katharyne Lescailje, and Lucretia Wilhelmina van Merken.³⁴ We grouped plays by author to minimise the effect of the author’s stylistic fingerprint.

A first step to evaluate the performance of gender is to compare the visibility and stage presence of male and female characters. We first measured the visibility of male and female characters by computing the percentage of male and female characters and the relative number of lines per gender group on the level of the play. Secondly, we computed visibility on the character level, based on the relative number of lines per character and the relative number of scenes in which each character speaks. All indicators can be normalised on a scale from 0.0 to 1.0, which enables comparisons between plays and characters from different plays. The high diversity and breadth of difference between main and side characters means the average values are of little use. Instead, we visualise the distribution of the data on the play level (fig. 2A, 2B) and the character level (fig. 3A, 3B) as violin plots. The horizontal bars in the belly of the violins represent the average proportion of male and female characters and lines in all plays.

Fig. 2A, 2B specify the dominance of male characters in the corpus. The uneven distributions visualised by the violin plots indicate that in most plays, male characters outnumber the female. In thirty-nine plays (40 percent), less than 25 percent of all characters are female, whereas there are no plays with less than 25 percent male characters. This imbalance in stage presence between the gender groups correlates with an imbalance in the discourse on stage: in eighty-one plays (82.7 percent), more than 50 percent of all lines are uttered by male characters.

A closer look at the visibility of the gender groups on the character level (fig. 3A, 3B) reveals that female characters are not necessarily likely to be less prominent or less visible on stage than male characters. The relatively similar violin plots indicate that an average female character speaks more lines and appears in more scenes than an average male character, despite the lower stage presence of female characters overall. In other words: female characters are a minority on the early modern Dutch stage, but when they do appear, they generally tend to play larger roles than most of their male counterparts.

Besides visibility, gender differences can be measured in terms of the narrative position of male and female characters. Following previous applications of network theory to narratological analyses of drama, we operationalised narrative position as the centrality in the social network of the play.³⁵ We constructed relationships (edges) between two characters (nodes) when they appear in the same scene and if they both speak at least one line during that scene. The higher the number of scenes in which they both speak, the stronger (the weight of) their relationship. We can construct interaction networks of single plays if we accumulate all those interactions for all possible character pairs. Characters who play a key role in the story world – i.e., who interact with many other characters – end up in the heart of those networks. Based on these interaction networks, we can compute the normalised centrality of each character as the number of edges per character compared to all characters in the play. There are more advanced metrics for centrality such as betweenness or eigenvector centrality to operationalise an individual’s ‘power’ or ‘influence’ on the network, but relative degree centrality suffices as a straightforward indication of a character’s position in the social world of the play.³⁶

A comparison between the centrality scores of male and female characters shows that the narrative position of the two gender groups is quite similar. Fig. 4 visualises the distribution of degree centrality scores, which is similar in each gender group. The average centrality of female characters (0.60) is slightly higher than male characters (0.54), but given the large variance of centrality values in both gender groups (indicated by the oblong-like shape of both violins), the difference proves marginal. This means that on average, female characters are not more or less central than their male counterparts in the social worlds of early modern Dutch plays.

However, as a result of the overall imbalance between male and female characters in most plays, the range of interactions is not equal between the gender groups. The polyphony that is so fundamental to theatre is in many cases limited to dialogues between men: of all the edges constructed in the social worlds of the plays, 1937 represented relationships between two male characters; 1551 of all the edges concerned a relationship between a male and a female character. Only 358 edges were female to female. Whereas these numbers should not surprise us, given the high number of male characters in the corpus, it is good to reflect on the consequences of these numbers for the question at stake in this article. When male characters dominate the interactions and therefore the discourse on stage, it is possible that dramatic discourse in general has a tendency towards ‘masculine’ (socio-)linguistic norms. In the next sections, we apply various forms of computational text analysis to evaluate the existence and qualities of this supposed binary opposition between ‘masculine’ versus ‘feminine’ language in character speech.

In our first experiment, the lexical approach, we tested whether characters tend to have gendered lexical preferences. Such preferences could be revealed by socio-linguistic vocabularies that are characteristic of their gender group. To measure the existence of these preferences, we ran several experiments with different settings, which are listed in tab. 3. In each experiment, character speech was fragmented into several ‘chunks’ of text. The column chunk size (words) refers to the length of the speech fragments that are used. This length is connected to not only the sample size (gender) which refers to the size of the number of male and female speech fragments, but also to the test set size (gender), which refers to the number of male and female speech fragments on which the model was evaluated. We experimented with the n-gram range: a range of (1,1) means that the model only looks at unigrams, in other words, single tokens. A range of (1,2) means that both single tokens and bi-grams (combinations of two words) are taken into account, and a range of (2,3) means that only bi-grams and tri-grams are taken into account. We have also run experiments without lemmatising the data, but since their performance was slightly lower than the experiments in which we did use lemmatisation, we only continued with experiments on lemmatised speech texts.

Applications of machine learning like the one at stake here are usually evaluated with the so-called F1-score, a number representing the proportion of correct classifications compared to all cases that had to be classified (‘recall’) and to all cases that were classified, correctly and incorrectly (‘precision’). In our case, the F1-scores per gender indicate how well the model is able to predict the gender of a certain character’s speech. We report the average F1-scores over fifty iterations. Overall, we see that the F1-scores increase when we increase the chunk size, suggesting that larger text fragments provide more reliable gender cues. Including bi-grams or tri-grams has no substantial effect on the models’ performances: for a chunk size of four hundred, the best results are obtained with unigrams, while for chunk sizes of three hundred and five hundred, bi-grams and tri-grams perform slightly better. However, the differences are minimal, with all scores falling between 0.726 and 0.757. While the performance improves with longer chunks, the model still makes a notable number of misclassifications.

Besides specific gendered vocabularies, we tested whether characters speak in gendered semantic fields or ‘discourses’ – using language in a specific context and sense that is distinguishable as either typical for male or typical for female characters. We considered this experiment the ‘semantic’ approach. In our experiments with the models GysBERT-v2 and m-e5-large, we use the same chunk sizes as in our previous lexical experiments (tab. 3). The F1-scores per gender are included in tab. 4.

Here, we see that the scores remain stable across chunk sizes. With scores between 0.735 and 0.756, they fall in the same range as the performances with the lexical models. For both embedding models, the F1-scores are nearly identical across genders, suggesting a balanced performance. The GysBERT-v2 model slightly outperforms the m-e5-large model, but the differences are small. While these results indicate that all models capture some gendered semantic patterns, they are not strong or consistent enough to enable highly accurate classification.

In addition to gendered vocabularies and gendered discourses, we searched for traces of gendered ‘styles’ in character speech. To assess the existence of a gender binary in the style of character speech, we performed a cluster analysis of all male and female character speech extracted from a selection of plays written by one single playwright. This analysis constructs dendrograms (or ‘kinship trees’) visualising clusters of stylistically similar documents based on the closeness (or ‘kinship’) between all possible document pairs in the selected sample. In our case, a document represented all speech by all male or by all female characters in a given play. This means that each edition is split into two documents: one for male character speech, one for female. We thus questioned whether the clustering of character speech could be explained by gender.

The dendrograms in fig. 5A, 5B represent the stylistic clusters of all-male and all-female speech in the four sub-corpora, where male speech is represented by a grey/gray colour label, and female by a black label. The trees represent a simplification of all document-relationships in the corpus. They can be read like a family tree. The position of each sub-branch on the horizontal scale indicates the stylistic closeness (or ‘kinship’) of the (clusters of) documents; ‘siblings’ are genetically more closely related than ‘second cousins’, for example. The question at stake is whether characters are more closely related – in terms of their style – to other characters with the same gender identity from other plays by the same playwright. If this were the case, we would expect to find coloured clusters in the dendrogram, with a clear separation between ‘grey/gray’ and ‘black’ documents. This pattern does not occur in the oeuvres by Krul, Lescailje, and Vondel, even if we repeat the analysis with different parameters, using smaller and larger selections from the vocabulary (with 50, 150, 200, and 250 mfw), or bigrams and trigrams instead of single words. The conclusion here is that gender cannot explain the stylistic clustering in these selections of plays: there is no correlation between character gender and speech style. Van Merken’s oeuvre, however, seems to be an interesting exception, although the pattern does not always reappear with different parameters. Still, this analysis provides some evidence of a ‘typical’ male or female style in Van Merken’s characters.

Finally, we used the oppose-function in Stylo to determine the strongest lexical markers of the distinction between male and female speech. This function computes Zeta-scores for all words appearing in the two text groups (male versus female speech). The Zeta metric is essentially a normalised indication of the deviation from the average frequency of a single word in all the tested documents. If, for example, the frequency of the word ‘gun’ in male speech is higher than one would expect given its frequency in the speech of an average character, then this would translate into a higher Zeta score. Fig. 6 shows the words with a high negative Zeta score (‘Avoided’) and a high positive Zeta score (‘Preferred’) in female speech compared to male speech. The graph shows all words in the vocabulary with a Zeta score higher than 0.05. The higher the score, higher on the list, the higher the predictive value of each word.

By instructing the program to determine the strongest indications of ‘maleness’ and ‘femaleness’, we are asking it to find the most stereotypical differences between male and female speech. Therefore, we should not be surprised that it gives us stereotypes in response. The most predictive words for female characters are words related to emotional descriptions (verdriet, gemoed, smart, rouw, hart), love (liefde, lief, lust), emotional exclamations (ach, helaas, ei, och), and family (moeder, vader). Typical male words are words related to politics (volk, land, rijk, troon, koning) and warfare (heir, kracht, geweer, wapen). These results confirm that some registers are indeed gendered. Moreover, they might in these cases help us to identify gendered dramatic functions. The fact that female characters are more likely to do the emotional labour in the play (indicated by the high frequency of emotion words in female speech), might lead to the interpretation that female characters often function as the mirrors reflecting the emotions expressed on stage by making those emotions explicit. However, we should not forget that fig. 6 highlights only a small sample of the vocabulary, representing the extreme differences between the two gender groups, while obscuring the large proportion of vocabulary that is apparently not gendered at all. It is a common fallacy in computational gender studies to ascribe too much meaning to observed gender differences based on small selections from the vocabulary, computed with metrics designed to find deviations and exceptions.³⁷ We prefer to read the small number of lexical markers of gendered speech that passed the 0.05-threshold as another confirmation of our assessment that most of the male vocabulary is not drastically different from most of the female vocabulary.

Our analysis confirms that male characters significantly outnumbered female characters on the early modern Dutch stage. Given the fact that most actors were male in the studied period, Dutch dramatic discourse was mostly spoken by male characters embodied by male actors. In terms of stage presence and visibility, women were a minority in the fictional worlds of the Dutch theatre, in proportions similar to the Spanish and English story worlds created by Lope de Vega and Shakespeare.³⁸ In terms of network centrality, however, female characters often hold a key position within their fictional social worlds, which seems to underline their relative exceptionality: female characters appear less on stage, but when they do, they tend to play important roles.

We questioned the function of character speech in the performance of early modern gender roles within this theatrical reality dominated by men. We tested the distinction between male and female character speech on three levels (lexical, semantic, and stylistic) and found only weak evidence of this distinction, except in a few stereotypical registers. Our lexical and semantic approach demonstrates that machine learning and large language models can be used to make an educated guess about the gender based on character speech. However, the performance of these models is relatively poor, especially compared to similar studies applying llms for gender classification in English or contemporary literary texts. Note that our models perform better than a random classification, which would lead to F1-scores around 0.50. Nevertheless, 20 to 25 percent of the cases are still misclassified. We interpret this inability to convincingly classify gender using computational methods as an indication that the binary categories we are modelling (male versus female) do not operate in such a binary manner in the texts under scrutiny, early modern Dutch theatre editions. If computational methods fail to capture gender distinctions in character speech in 20 to 25 percent of the cases, then either the methods are unfit for this task, or these distinctions are only partially present in the texts. Since similar studies have shown these methods to be fit for purpose when applied to modern English novels or modern Dutch novels, we opt for the latter explanation: there is no such thing as a typical ‘male’ or ‘female’ vocabulary or style in early modern Dutch character speech.³⁹ Returning to our hypothesis about the possible effects of the fluidity of gender in Dutch drama on the performance of gender, we conclude that character speech was not specifically gendered.

Before we turn to an interpretation of this result, we need to consider four characteristics of our corpus that might complicate comparisons to gender binaries in other literary corpora. First, the corpus is relatively small and the variation in spelling and morphology relatively large compared to the corpora analysed in similar studies. It is possible that gender distinctions become more pronounced when using larger samples. Secondly, the form of early modern Dutch theatre is relatively conventional compared to, for example, contemporary novels: all texts from the corpus are written in (paired) verse, and usually following a strict (iambic) meter. This also means that lines spoken by different characters often rhyme. The formal conventions of Dutch drama thus complicate the emergence of gender-specific vocabularies and styles. Thirdly, dramatic discourse is by definition limited to character speech, which is less suitable for gendered character descriptions or perceptions by narrators and focalisers, for example. Finally, we already suggested above that the absence of gendered speech might be explained by the overrepresentation of male characters in the discourses on stage, and by extension in the predominantly male group of poets who were responsible for writing a large majority of the plays published and staged in the Republic. The absence of a gender divide in character speech could thus be a symptom of a skewed gender divide in the authorial styles represented in the corpus.

Notwithstanding these caveats, the conclusion holds that we found only weak evidence of a clear distinction between male and female speech from early modern Dutch plays. Overall, male and female characters did not speak in fundamentally different styles or about different topics. What could we infer from this fact about the role of gender in Dutch drama and in the Dutch Republic at large? The absence of a gender binary reminds us first of all that early modern gender roles were performed dialogically: characters were shaped by what they said as much as by what others said about them. The form of early modern plays thus enabled representations of individuality, identity, and gender that were constructed through interactions between characters. Moreover, these results indicate that female characters were generally not ‘othered’ in their male-dominated fictional world, at least not by their contributions to the discourse on stage. Stylistically and lexically, female characters were equal participants in the conversations on stage. They were not excluded from the discourses in which they spoke based on their gender (if they spoke at all, that is). This conversational equality on stage may have been a reflection of the many domains where men and women interacted and negotiated their place in the relatively egalitarian public spaces of the Republic: the street, the market, the café, or the horse-drawn barge.⁴⁰

We are inclined to interpret the non-existence of male and female speaking styles as part of a theatrical reality in which gender was unstable and complex due to the habit of cross-dressing – a reality that became even more complex after the first female actors entered the stage in 1655. Gender must have been marked by narratological, visual, and auditory features such as roles (the queen), costumes (her dress), and tone of voice (falsetto), but in terms of their speech text, male characters were not radically different from female characters. This fluidity may have been typical to the theatrical culture of the seventeenth century only, given our finding that character speech was in fact more clearly gendered in the late eighteenth-century plays by Lucretia Wilhelmina van Merken. Previous scholarly readers of Van Merken’s work noted the remarkable role of her female characters, as well as her deliberate foregrounding of female roles in the stories she dramatised.⁴¹ Our stylometric approach suggests that she also used distinctive stylistic features in her characterisations of female figures.

In Dutch literary history, distinctions between ‘maleness’ and ‘femaleness’ thus seem to have a history of their own, similar to the way in which Ted Underwood and others have shown that gender demarcations in modern English novels changed – becoming less pronounced, in their case – in the nineteenth and twentieth centuries.⁴² Adding to the studies by Pipkin, Van Elk, and Ferket on the representation of sexual violence against female characters, and the representation of femininity in Dutch plays and female discourses on stage, our study reveals that from a comparativist perspective, male and female characters are surprisingly similar in the ways they spoke onstage. Instead of contradicting previous analyses of gendered stage roles, substantiated by qualitative textual evidence, the computational lens enriches the historiography of early modern gender performativity by showing that male and female characters communicated in similar, perhaps even equal ways, performing their gender roles through dialogue – again, if female characters appeared on stage at all.

Finally, we should not forget that the fluidity of gender representations on the early modern stage sharply contrasted with the very real legal and social norms and boundaries that distinguished the position of men and women in early modern Dutch society. The symbolic domain of the theatre – where gender roles could be complex and fluid – thus challenged the everyday world around it – where those roles were clearly separated. This contract opens up new questions about the long history of gender boundaries, both as representations in the symbolic domain of literature and as a social, legal, and material reality outside the theatres. Developing larger corpora and new computational approaches will remain crucial in finding answers to those comparativist questions about the longue durée in the history of gender.