Gender Depiction in Portuguese. Distant Reading Brazilian and Portuguese Literature

1.1 Gender in Literature

The theme of gender roles in fiction texts has received increasing attention in the Digital Humanities community, as the following works testify.

Looking at English literature (104,000 works, from 1703 to 2009), Underwood et al. (2018) found that the gender difference between characters became less pronounced from the middle of the nineteenth century to the present day: Actions and attributes of characters became less defined by gender categories. In other words, gender roles tend to become more flexible. At the same time, they also found a decrease in the number of feminine characters, with the volume of fiction written by women from 1850 to 1950 dropping by half.

Exploring the Black Drama collection, which contains plays written between 1950 and 2006, Argamon et al. (2009) report poor results when trying to automatically distinguish the gender of the authors and/or characters. However, they found differences in the way masculine and feminine authors and characters use language. Feminine playwrights allocate more than half (52.1%) of speeches to feminine characters, while 34.7% of the speeches in plays by masculine authors belong to feminine characters.

Working with present-day Dutch literary fiction (170 novels published in one sample year), Smeets (2021) found the same imbalance between masculine and feminine characters. However, the author questions what he describes as a “perhaps naive mimetic assumption” according to which the relative absence of feminine characters is a result of their unequal status in society. From the results of his investigation, feminine characters, although fewer in number, occupy a relatively central position in their fictional social networks –– they display more relations, both more relations in general and more relations with important characters.

Hoyle et al. (2019), using 3,5 Mio. digitised books in English, analyses the lexical choices (adjectives and verbs) associated with feminine gendered nouns and found that positive adjectives used to describe women were more often related to their bodies than adjectives used to describe men. Following the same trend, Schulz and Bahník (2019) explores the depiction of male and female characters using the Google Books Ngram corpus, focusing on twentieth-century English-language fiction. The study analyses adjective-noun bigrams associated with the words man, woman, boy, and girl, and reports that adjectives associated with men are more positive (“honest”, “wise”, “honorable”, and “able”) than those associated with women (“vulgar”, “foolish”). As for preferences, “charming”, “fashionable”, and “warm” were relatively feminine words, while “lazy” and “mean” were relatively masculine words. On the one hand, men were described in decreasingly masculine terms throughout the beginning and end of the twentieth century; on the other hand, the masculinity of adjectives used to describe women started to slightly increase from 1968 to 2000.

Weingart and Jorgensen (2013) performed a computational analysis of gendered bodies in ca. 200 European fairy tales (German, French and Italian folklore texts translated into English). They show that feminine characters are more likely than masculine characters to be described with appearance-evaluative words, suggesting that men are associated with the mind and women with the body.

Cermáková and Mahlberg (2022) explore linguistic descriptions of gendered body language and compare nineteenth century British children’s literature (ChiLit Corpus) with contemporary fiction for children (the OCC2000+ Corpus, a subcorpus of the Oxford Children’s Corpus). Using a corpus linguistic approach, the authors study sequences of five words which contain at least one body part noun and a marker of gender. They found fewer clusters for feminine characters in the nineteenth century. The contemporary data suggest a trend for feminine and masculine clusters to become more similar, and an increasing range of options for the description of feminine characters and their interactional spaces. Using the same ChiLit Corpus, Cermáková and Mahlberg (2021) focused on nouns –– excluding proper names –– frequently used to label people, and found that Mothers are the most frequent occurring feminine character in the corpus.

It is also worth noting the existence of studies such as Cao and Daumé (2021) and Lucy and Bamman (2021). The first one explores the consequences of gender bias for machine learning. The paper investigates how different aspects of linguistic notions of gender impact an annotator’s judgements of anaphora, and points out that a significant possible source of bias comes from the annotations themselves –– from underspecified annotation guidelines and the human annotators. The authors emphasise that both, humans and systems, should not over-rely on cues such as names, semantically gendered nouns, and terms of address, relying on “relatively safe” cues like syntax instead. At the other pole of the machine learning approach, the study conducted by Lucy and Bamman (2021) raises questions on how to avoid unintended social biases when using large language models for storytelling. Focusing on how GPT-3 may perceive a character’s gender based on textual features such as personal pronouns (he/she/her, etc.), the work finds that stories generated by GPT-3 place masculine and feminine characters in different topics and exhibit many gender stereotypes: For example, feminine characters are more associated with family and appearance than masculine characters.

In this paper, we also try to contribute to the investigation of gender roles using works written in Portuguese. As a crossover between Corpus Linguistics and Digital Humanities, we use morpho-syntactic and semantic information automatically provided by the PALAVRAS parser (Bick 2014), and add extra semantic annotations, which are described below.

With Larson (2017), we recognise that using gender as a variable in Natural Language Processing is an ethical issue and that we need to explicitly explain what “gender” means in this work. As Larson points out, there are many views of how gender functions as a social construct. In this study, we treat gender as binary, since in the vast majority of works in our corpus, gender was mainly constructed in terms of the binary distinction femininity/masculinity. We acknowledge, however, that the category “gender” can be more complex than this binary distinction, and that these kinds of studies, which describe the cultural apparatus around gender for an extended period of time, do not in any way purport to assert what gender is, but only how it has been/is perceived. So they should not be used for reinforcing gender stereotypes, as warned against by Mandell (2019).

1.2 Previous Work for Portuguese

For distant reading of Portuguese, we are aware of some works dealing with characters in literature (Santos and Freitas 2019), as well as of the DIP challenge for automatic character identification in Portuguese (Santos et al. 2022b), to which we will come back later.

Our point of departure is the work by Freitas et al. (2022)¹ – and later extended in Silva’s master thesis (Silva 2021) – who have suggested a fourfold classification for human characterisation. Human attributes were organised in social, appearance, character, and emotional characteristics.

Using OBras, a corpus of Brazilian literature in the public domain (Santos et al. 2018), they studied 223 works by 25 Brazilian authors, two of them women (authoring 3 novels altogether), and observed the following trends:

• Men were more frequently described than women (60%-40%), which may be related to the fact that there were roughly more masculine characters than feminine ones in the same proportion.

• The most frequent masculine characterising words were bom (good), sério (‘honest’), rico (‘rich’), and alto (‘tall’), while bonita (‘beautiful’) was by far the top characteristic for women.

• Almost 50% of women depicting words were about beauty (namely bonita and bela).

• Character and social predication were most frequent for men; for women, social characterisation is reduced to married and rich.

• Emotional characterisations like feliz (‘happy’) were (almost) exclusively used for women.

We wanted to check whether these observations held true for a wider collection, including Portuguese literature as well.

1.3 A Brief Comparison with DIP

It is useful to compare and contrast our study with the recent DIP challenge for Portuguese (Desafio de Identificação de Personagens), an evaluation contest for identifying literary characters and some information about them in Brazilian and Portuguese works (Santos et al. 2022a, 2023). By describing them and pointing out the differences, we shed some light on different ways of looking at (roughly) the same data.

For DIP, the unit is the literary character, and so the challenge looked at their gender, their profession, occupation and/or social status, and their family relationships with other characters. In addition, “literary character” in DIP does not include all people.² In the present study, we try to look at all mentions of characterisation of people in the works, so all numbers reported in this paper are not per character, but per mention of people.

We will discuss and compare the findings about character gender in subsection 4.7.

1.4 The Importance of Studying Literature in Portuguese

Portuguese has a rich literary tradition, but unfortunately the digitisation efforts are lagging behind other languages. This has, for example, been discussed in Schöch et al. (2021).

Also, major actors in the big data landscape, no matter the high number of Portuguese speakers in the world, have not endowed Portuguese with the “current” tools that are available for other languages, even with much fewer speakers/readers/writers, like Hebrew or Italian: There is, for example, no Google Book N-grams³ service for Portuguese.

Likewise, recent reviews of the computational literature landscape, because they do not find enough internationally published DH papers on Portuguese, have decided not to review or include papers in Portuguese, therefore contributing actively to the lack of information on Lusophone materials and studies. For example, Schöch et al. (2022, 4) state:

“several languages, however, were represented only with relatively low numbers of articles or papers, and in order not to misrepresent the research communities these publications stem from, we decided not to take the materials in several languages into account […].”

This is one of the reasons why we are writing this paper for an international audience. Maybe the results are not so different than the ones our English-speaking or English-studying colleagues obtained, but they are novel because they are obtained from completely different data.

2.1 Corpus

We used Literateca version 11.1, created on 26 May 2023, comprising ca. 32 Mio. tokens of (original) prose (excluding drama) from 1700 on. A quantitative overview of the material is in Table 1. Figure 1 shows the distribution of the material in time, by size in words.

Figure 1: Distribution of words per decade.

Literateca is the result of the merging of several literary corpora written in Portuguese, and thus has some particularities:

• It includes literary works by canonical authors, but also other works by those canonical writers that are not usually or necessarily deemed literary, such as newspaper chronicles, letters, memoirs, and even scholarly works such as history books or ethnographic studies, and travelogues. For earlier centuries, even sermons are included. However, these genres are only included for canonical writers.⁵

• It includes drama, poetry, and prose.

• Some of the works have updated orthography, others keep the original orthography. Given that there have been several norms of Portuguese spelling across the centuries, this means that there can be a variety of forms for the same word.

• While some authors have all their works included, others have only a few, or just one. Especially for non-canonical writers, there is no claim to completeness.

• Given that the works have been digitised by different bodies and with different tools and for different purposes, there is no claim to homogeneity: Works can come from the first or the last paper version, they may keep their prefaces or not, they have different ways of describing chapters, etc.

• All works are marked with author, author gender, date of publication, variety of Portuguese, genre, and whether they are original or translated. Some texts are also classified by the literary school they belong to.

We tried to use as much of this material as we could, but we removed poetry and drama. Poetry is probably a natural choice to remove because of syntactic idiosyncrasies – and therefore a worse parser performance –, and because we believe that poetry has not so many mentions of fictional characters. We removed drama, also in prose, because it was heavily unbalanced, given that most of the plays were from Portugal.

As to prose, we started to use everything published since 1700. It is, anyway, important to recognise that we do not have a balanced corpus, and the lion’s share is fiction. We then selected different subsets for different research questions.

• Just fiction and just non-fiction, to see whether the character depiction was different across the fiction divide.

• Just works published after 1840 to compare Brazilian and Portuguese authors.

• Just fiction published after 1840 to compare Brazilian and Portuguese literature.

See Figure 2 and Figure 3 for a bird’s eye view of the genre distributions in total and in fiction.

Figure 2: Genre in the full corpus. The unit is the work. In red are the works written by Portuguese authors.

Figure 3: Genre in the fiction corpus. The unit is the work. In red are the works written by feminine authors.

Only in Figure 3, we include the variable author gender, since it is only in fiction that we have texts written by women.

In Table 2, we give the number of words in the material published after 1840.

2.2 Gender Attribution

We explore the influence of gender in both character description and authorship. Masculine and feminine gender labels were manually ascribed to writers, as our corpus contains works written by canonical authors, which are widely discussed in literary studies. For the non-canonical authors, gender was attributed either based on adjective/inflected forms used in prefaces or based on their proper names. As for the characters, the gender labels were automatically assigned by the PALAVRAS parser, and then manually revised by linguists (Rocha et al. 2019; Silva 2021). The linguistic clues that were followed on attributing and revising gender were syntactic agreement and morphological features.

Portuguese is a Romance language that forces the speakers to specify the gender of nouns (both common and proper nouns) and adjectives. The main formal clue to distinguish masculine and feminine forms is the word’s ending: Masculine forms tend to end in -o, feminine ones tend to end in -a, and those ending in -e can be both feminine and masculine – ponte (‘bridge’) is feminine, and pente (‘comb’) is masculine. However, there is no perfect equivalence between the ending in -o or -a and the masculine or feminine gender, respectively – planeta (‘planet’) is masculine, and tribo (‘tribe’) is feminine. Therefore, observing syntactic agreement between the head noun and its modifiers is the most reliable way to assign morphological gender.

When calculating the gender of depicting words, we take into account the gender of the nominal head (noun, proper noun or pronoun) being characterised, not the gender of the words (modifiers) associated with it. This choice is due to the fact that, although adjectives can be inflected for gender in most cases, the search patterns we used also retrieve nouns, which do not admit inflection. Thus, nouns like anjo (‘angel’) will always be masculine, even if the mentioned angels are feminine. When considering the gender of nominal heads, anjo, although a masculine common noun, is classified as a feminine classifier if it modifies a feminine character.

4.1 Does Textual Genre Matter?

Does it make more sense to look only at literary texts, removing travelogues, essays, history and political writings?

On the one hand, we kept all the material because we wanted to look at the way people described people in Portuguese, but then it is also conceivable that the kinds of information about people are rather different when you write the history of the Inquisition, an essay about your fellow writers, or a report of you crossing Africa, compared with a narrative in which you introduce fictional characters.

So, we reproduced our queries, removing all texts not classified as novels, novellas or short stories (see the new numbers in Table 6).

It is interesting to see that removing the non-fictional prose genres does not change the relative order of the subcategories but increases the percentage of feminine characters, from 30.0% to 32.4%, and characterised feminine characters, from 33.2% to 35.0%.

As to the characterisation of masculine and feminine characters, we have similar trends to those presented for the full material, as shown in Figure 7: Masculine targets are characterised, by far, by their character, while feminine targets are (almost) equally characterised by their appearance and their character.

Figure 7: Relative characterisation per gender in novels, novellas, and short stories.

For the non-fiction part, let us see whether the picture is different. In Table 7, we describe the masculine and feminine characterisations in the (considerably smaller) non-fiction part.

The percentage of feminine characters and feminine characterisations shrunk considerably to 16% and 18%, confirming that women are even less important in the public sphere.

We see that social characteristics are – globally – more frequent than appearance. ‘Character’ remains the most important form of describing people, and ‘emotion’ the least.

In Figure 8, we present the distribution of the four kinds of features and see that the few mentions of women that are present have a large proportion of appearance descriptions, even more in non-fiction than in fiction.

Figure 8: Relative characterisation per gender in non-fiction.

4.2 Differences when Describing Masculine and Feminine Characters

The previous figures show that ‘appearance’ is more frequently used when describing feminine characters. Based on the entire data set, this can also be seen in the bar plot in Figure 9.

Figure 9: Relative characterisation per gender in the whole material as a bar plot.

However, this is just the tip of the iceberg. The analysis of depictive words preferentially used with masculine and feminine characters can be more revealing than the general analysis we presented in Figure 9, which takes into account the whole range of depictive words. In order to be evaluated as ‘preferred’, a word must (i) be used for masculine targets at least 80% of the occurrences, or for feminine targets more than 60% of the occurrences; and (ii) have a total frequency of 4 or more.

In cases where different lexical items correspond to gendered male/female pairs (mãe/pai (‘mother/father’); rainha/rei (‘queen/king’); namorada/namorado (‘girlfriend/boyfriend’) etc.), we manually grouped the elements of the pair as if they shared the same lemma, so that they could be included in the preference count.

The new data are presented in Figure 10, which shows a slightly different picture, in which (i) words of the emotional axis are almost not seen at all and almost disappear with the feminine characters, (ii) the balance between ‘appearance’ and ‘character’ in feminine depiction gives way to a characterisation based mainly on ‘appearance’, which accounts for half of all preferred feminine characterisations, and (iii) ‘appearance´, the second most frequent characterisation (of both masculine and feminine characters), drops to third place when associated with masculine characters, and rises to the first place when associated with feminine characters.

Figure 10: Preferred characterisation per gender.

The ‘appearance’ axis has a raw frequency almost similar for both genders, but Figure 11 and Figure 12, complementary to Figure 10, provide a few details that enrich the analyses.¹⁵

Figure 11: Preferred characterisation of masculine characters.

Figure 12: Preferred characterisation of feminine characters.

As noted in previous studies, typically feminine social characterisations relate to the family environment (mãe (‘mother’), prima (‘cousin’)). However, mentions of the marital status are the highlight: (casada (‘married’) and viúva (‘widow’) are the most frequent words, but adúltera (‘adulteress’) is frequent as well. Conversely, marital status is absent as typical masculine social characterisation. They are rather related to (positive) social recognition such as ilustre (‘illustrious’), célebre (‘famous’), notável (‘remarkable’), famoso (another word for ‘famous’), and poderoso (‘powerful’).

On the feminine emotional axis, words associated with love and sweetness (adorada (‘adored’) and meiga (‘sweet’)) stand out, but also words associated with sadness and insecurity (pobre (‘poor’), chorosa (‘tearful’), ciumenta (‘jealous’), ofendida (‘offended’)), and fear (espavorida (‘terrified’)). On the other hand, bravery is the masculine highlight: valente (‘brave’) is, by far, the most frequent word and atrevido (‘cheeky/audacious’) is in the sixth place. Humility (humilde (‘humble’)) and anger (furioso (‘furious’)) rank second and third, respectively. Anxiety also appears: desesperado (‘desperate’) is the fourth most frequent emotional word for masculine characters.

Finally, masculine characters seem to be taken by surprise more often than feminine ones, frequently being assombrado (‘haunted’), surpreso (‘surprised’), and maravilhado (‘marveled’), which might be due to their role in narrative events.

‘Appearance’, although highly typical for feminine targets, varies relatively little in terms of the most frequently mentioned attributes: Beauty (bonita, formosa, bela, linda, Portuguese words for ‘beautiful’; encantadora (‘charming’)) or the lack of it (feia (‘ugly’)) are the most frequent features. In the masculine appearance axis, age and size, instead of beauty, are the most frequently mentioned attributes (velho (‘old’) and jovem (‘young’); robusto (‘robust’), grande (‘big’) and baixo (‘short’)).

On the typically masculine character axis, positive traits such as grande (‘great’), simples (‘simple’), verdadeiro (‘real’), valente (‘brave’), livre (‘free’), and hábil (‘skillful’) stand out. Other highly mentioned positive traits are generoso (‘generous’) and habilidoso (‘skillful’). Negative highlights are mau (‘bad’), terrível (terrible), and rude (‘rude’). For the feminine targets, the highlights are, in general, positive and associated with virtue: virtuosa (‘virtuous’), santa (‘holy’), and inocente (‘innocent’). Other typically feminine characterisation words are meiga (‘sweet’) and dócil (‘docil’), but we also see fraca (‘weak’), which contrasts with masculine strength.

4.3 Does the Gender of the Author Matter?

Do these findings vary according to the author’s gender? In our material, see Table 8, feminine authors use more appearance descriptions than masculine ones, as shown in Figure 13.

Figure 13: Characterisation by masculine and feminine authors. Note the different sizes in the y-axis.

However, there is a huge difference in the size of the compared material: There are only 1.2 Mio. words written by women compared to almost 32 Mio. words written by men. In fact, this is an inescapable problem, given the reduced number of texts by women in our corpus: only 19 authors who wrote 33 works in prose. ¹⁶

Even though the material is very unbalanced, we tried to discern any interesting trends in the works written by women in terms of whose appearance is described more – could it be that they would emphasise or concentrate more on the appearance of masculine characters?

We get 265 appearance descriptions of feminine characters and 319 of masculine characters in 985 characterisations of feminine characters and 1,195 characterisations of masculine characters. In other words, 26.9% of feminine characterisations and 26.7% of masculine characterisations involve their appearance. But we acknowledge that the numbers are too small to be conclusive. In any case, it is conspicuous that both genders have roughly the same characterisation frequency in literature written by women.

Despite the imbalanced data, Figure 14 shows a preferential characterisation of both characters and writers in terms of gender. Below, we sketch some differences between human depiction in works written by men and women. The main difference is the increase of ‘appearance’ in masculine characterisation in works written by woman.

Figure 14: Preferred characterisation by masculine and feminine authors.

Beginning with feminine characters and focusing on women writers only, we found that married is no longer among the most frequent social depictions, but widow and single remain. Despite still being frequent, less space is devoted to beauty in works written by women. By contrast, age is more present: young and old. As for emotional characterisation, happy and adorable are the highlights, and none of the preferred emotional words relate to sadness. As for character, the highlights of feminine depiction words are honest, infamous, crazy, refined, and dangerous. In the social axis, masculine characters are mainly married and noble. Positive emotions are present for masculine characters as well (like happy/pleased, enthusiastic), but bravery (brave) has only one occurrence. Masculine ‘appearance’ follows the general trend, and masculine characters are mainly kind and honourable.

4.4 Differences between Brazil and Portugal

Are there differences between the two countries with regard to people’s characterisation?

We compared the works from 1840 to the present day (Brazil became independent in 1822, and, as already mentioned, for the 1830 decade we only have one work by a Portuguese author).

We decided to compare only novels, novellas and short stories between the two countries because the non-fiction parts differ widely: While we have a large body of texts on history on the Portuguese side, we have almost only short essays in newspapers on the Brazilian side. The results are presented in Table 9.

We can see that the numbers of ‘character’ and ‘social’ characterisation are somewhat higher in Portuguese literature, while the other categories – especially emotion – are more pronounced in Brazilian literature. One may wonder whether this is due to a more socially rigid society in Portugal, or whether the cause lies in the historical novels (almost absent in the Brazilian material and quite frequent in the Portuguese material).

We also investigated whether the differences among genders are more obvious in the Brazilian material or different from the ones in the Portuguese material. For this, we created Table 10, where we can see that Brazilian literature has a higher proportion of mentions of feminine characters (36.5%) than the Portuguese (29.7%). This may again be due to the historical novels, but needs to be investigated further.

In Table 10, we see that the social status of male characters is more important in Portuguese literature.

If we now compare the distribution by country and by gender, presented in Figure 15, masculine characters seem to be similarly depicted, although in Portuguese-authored works there is a slightly more balanced distribution between appearance, social and emotion axes. In Brazilian-authored works, besides the emphasis on ‘appearance’, there is proportionally less use of the character axis, which leads to a smaller difference between characterisations by ‘appearance’ and by ‘character’. For feminine characters, there are relatively fewer mentions of their social status and emotional states in Brazilian-authored works.

Figure 15: Characterisation by country.

4.5 Differences among Authors

In Table 11, we show the distribution of the types of characterisation for 12 canonical authors, six Brazilian and six Portuguese.

We can see that there are some differences among these authors. They agree in that none of them emphasises an explicitly emotional description, and several authors follow the “general” pattern in fiction: first ‘character’, then ‘appearance’, then ‘social’, and finally ‘emotion’: Machado de Assis, Eça de Queirós, Aluísio de Azevedo, José de Alencar, Júlio Dinis, Teófilo Braga, and Alexandre Herculano.

However, in José Manuel de Macedo, Coelho Neto and Raul Brandão ‘appearance’ is the most frequent characterisation and ‘character’ is the second most frequent.

As to the relative order of ‘character’ and ‘social’ characterisation, Humberto de Campos is the only one who reverts the “canonical” order, using more ‘social’ characterisations than those reflecting ‘character’, while Camilo Castelo Branco (incidentally the author with the highest number of works in Literateca) is the only one who describes more ‘social’ than ‘appearance’.

In any case, there are also differences in the number of characterisations provided by each author: Figure 16 illustrates how much each author depicts, i.e. how many characterisations they use per number of words.

Figure 16: Characterisation by author. From left to right, Humberto de Campos, José Manuel de Macedo, Aluísio Azevedo, Camilo Castelo Branco, Júlio Dinis, José de Alencar, Eça de Queirós, Machado de Assis, Coelho Neto, Alexandre Herculano, Raul Brandão, and Teófilo Braga.

In Figure 17, we represent each author in a plane formed by internal and external characteristics.

Figure 17: Characterisation by author in terms of type and relative weight of characterisation.

4.6 The Influence of Literary School

For a subset of the works of Literateca, we have metadata about the literary school to which they belong, as described in Santos et al. (2020).

We selected all works marked as romantic in one group (11,850,395 words, 175 books) and those marked as realist or naturalistic (7,616,384 words, 121 different books) in another group¹⁷ to see whether one could identify differences regarding people’s depictions just based on this fourfold sub-classification, and also according to the gender of who gets characterised. The results are presented in Table 12 and in Figure 18.

Figure 18: Characterisation per literary school and per gender.

The first interesting remark is that there are (relatively) more mentions of feminine characters in realist works than in romantic. However, 10.9% of the feminine occurrences are characterised in romantic books (and 9.9% of masculine occurrences), but only 9.8% in realist ones (compared to 9.5% for masculine).

We see that in romanticism, there are far more ‘character’ characterisations of masculine characters than in realism, where the relationship across all kinds of characterisations is stable across genres. In addition, realism describes the physical appearance of both genders, while romanticism prefers feminine appearance.

4.7 Going back to DIP

DIP has clearly demonstrated that there are fewer feminine characters in Lusophone literature. In this study, however, we see that those feminine characters are relatively more characterised, at least for ‘appearance’, than the masculine ones.

Ideally, and for the near future, we would like to connect the two studies/activities/forms of distantly looking at literature and provide, for each literary work, not only their description in terms of characters (as DIP does) but also how each character is characterised, using the present work and some form of anaphoric resolution of the non-proper name depictions and of those cases where human subjects (whether or not proper names) are omitted (Freitas and Souza (2021) found omitted subjects in 41% of clauses in Brazilian literature material).

We might therefore link types of characters with particular clusters of properties, like the beautiful rich woman and the poor honest lad and the evil old priest.