The starting point of my first case study is that the poetry of early modern Dutch poets became increasingly embedded in a Renaissance and classical tradition. Inspired by the manner in which Italian and French writers adapted the traditions of Greek and Latin poetry, Dutch poets also appropriated their topics and genres. An important example of this embedding in the Renaissance tradition was the practice of the Petrarchan genre. The Italian Renaissance poet Francesco Petrarca – or Petrarch, as he is known in English – was a fourteenth-century Italian poet, whose fame extended beyond his country’s own borders already during his lifetime.¹⁰ One of the reasons for Petrarch’s fame was the popularity of his love poetry. His love sonnets were distinguished by the worship of a charming but unreachable woman, which meant that love always causes both happiness and pain in his poems. The physical beauty of the beloved is often described using natural metaphors and mythological references. The lover is often characterized as someone with a split personality, who struggles with suicidal tendencies.¹¹

In his study on seventeenth-century songbooks for youngsters, Grootes argues that one would not expect Petrarchism to become popular in the Low Countries, since the trope of the unreachable woman was paradoxical to the actual surplus of women in Amsterdam and the province of Holland at the time: ‘At first sight, there seems to be no ground for the humbly pleading of the lovers, nor for the presentation of the songbooks as instruments to make the girls a little more obliging.’¹² Even so, Grootes suggests that the ideology of Petrarchism, with its emphasis on restraint and the lack of sexual boldness, probably fitted perfectly with the marriage policy advocated by the parents.¹³ He also cites the literary scholar August Keersmaekers, who has argued that the songbook Den Nieuwen Lust-hof (1602) could be interpreted as part of ‘a culture of piety propagated and imposed from above’, in opposition to the older songbooks.¹⁴ The quantitative methods used in this study are ideally suited to explore whether the popularity of the classic genres, and the Petrarchan genre in particular, can be found in the current research corpus. Furthermore, qualitative research has not paid attention to the moment of decline in popularity of these genres, something which the methodology used in this study might also shed light.

Another claim that will be tested in this study is how songs were used to articulate a feeling of patriotism. As Cornelis van der Haven notes, a long tradition of aggressive patriotic rhetoric can be found in Dutch literature, a tradition that is strongly related to the Dutch liberation myth of the struggle against Habsburg Spain that would be constructed in later centuries.¹⁵ In times of war or political crisis, feelings of patriotism and unity are aroused and propagated by, among others, poems and songs.¹⁶ A social activity, singing together created a feeling of belonging to a particular group, and contributed to the formation and maintenance of a social identity and an imagined community.¹⁷ Regarding the early modern Netherlands, Dieuwke van der Poel, Louis Grijp, and Wim van Anrooij have distinguished seven singing groups. The ‘political group’ is most relevant in this context, which they describe as a group ‘(often with a specific religious affiliation) such as the anti-Catholic and anti-Spanish Geuzen (‘Beggars’)’.¹⁸ Their protest songs – known as beggar songs – became very popular in the late sixteenth century. Initially they were published and disseminated via broadsheets, but before long they were bundled together and published in a songbook called Nieu Geusen Liedtboexcken (The New Beggars’ Songbook). The oldest edition of this song book dates from 1577, and several reprints were published in subsequent years. The songs were intended as inflammatory news announcements that would encourage revolt.¹⁹

Songbooks were the most obvious way to articulate an identity, in other words, to express one’s belonging to a group. Often group identity was already prominent in the title, for example in the Nieu Geusen liedtboecxken.²⁰ Furthermore, identity was also articulated through song lyrics – which is something I will explore in this article. Although the Wilhelmus is the most well-known of all beggar songs, and one which retained both its political overtone and signalling function long after the 1648 Peace of Münster, there were many other songs.²¹ While their popularity probably did not reach the height of the current Dutch national anthem, their tropes of nation and identity likely resonated with the inhabitants of the Dutch Republic. That being said, there were also songs expressing love for the fatherland without specific political orientation, in which the whole Dutch nation was addressed as one political group, especially in times of war. During the seventeenth century, patriotic ideas became more widespread, a movement which was not only down to the beggars and their repertoire. Orthodox Protestant ministers also started to elaborate on the idea that God, the Netherlands, and the House of Orange were inextricably connected.²²

To test whether these trends can be found in a larger research corpus, I have relied on the Dutch Song Database, which contains the metadata of approximately 175,000 songs in the Dutch language, from the Middle Ages up to the twentieth century. The database was initiated in the early 1990s by Louis Grijp, who continued to lead its development until 2015. During these years, many research and documentation projects have been carried out, and so an enormous amount of high-quality data has been collected. Nowadays, the Dutch Song Database offers a rich cross-section of Dutch song culture throughout history.²³ It contains several genres, including love songs, satirical songs, psalms and other religious songs, folksongs, and children’s songs. For every song, the source for the text and/or the melody are indicated. Other available metadata include the author’s name (if known), the first line, the number of stanzas, genre, melody, stanza form, and number of verses. In addition, the results of the project Dutch Songs On Line became available in 2014.²⁴ The project has made 53,351 full song texts accessible online. Of these texts, 29,590 have had both their lyrics and metadata encoded with tei compliant xml, which provides the publication date, geographical location, melody, and genre. For the purposes of this study, I have limited this sample to songs that were printed and published between 1550 and 1750, which resulted in a corpus of 22,297 songs. This number does not yet include the reprints and appearances of these songs in other songbooks. Since I wish to examine which song topics were more prevalent than others, it is necessary to include these data as well. Although the number of different lyrics will remain the same, the distribution of the songs can change. Suppose, for example, that songs on politics and history, which are not very prominent in the corpus based on the data in tab. 1, are reprinted so often that their number increases massively, while songs on love are rarely reprinted. This would influence the distribution of topics in such a way that political and historical topics are more dominant than love topics. Expanding the corpus with reprints resulted in a corpus of 43,772 songs.²⁵ Their distribution over time is visualized in fig. 1.

Tab. 1 gives an overview of the distribution of songs over the various categories, as labelled in the Dutch Song Database.²⁶ The majority of the songs fall into the category ‘religion’, followed at a distance by ‘love and sex’ and ‘seasons and annual events’. One might ask why a study on topics in song lyrics is useful, given that data on categories already exist, but a closer look at tab. 1 shows the necessity of such a study. Not only does the categorical division seem rather arbitrary, there are also more than seven thousand songs that have not been assigned to a specific category at all. As such, the available data do not give an accurate overview of the distribution of songs over genres. Moreover, the wide-ranging categories do not reveal how subdivisions within a category are distributed. Are all religious songs on the same aspects of religion? Does ‘love and sex’ mean that some songs are on love, and others on sex? There is an important difference between a genre and a topic, since a genre can comprise many different topics. A song within the genre ‘love and sex’ might be a tragic love song, but it could also be a happy, uplifting love song. A quantitative analysis of the songs, using topic modelling, can therefore offer useful information: it offers a more fine-grained classification of the different genres, and will allow the topical fluctuations over time to be made visible.

Tab. 1 also provides information about the average song length within each genre. It shows that the average length of a song belonging to the genre ‘religion’ is about one hundred words longer than a song belonging to the category ‘love and sex’. The most outstanding outlier is the genre ‘politics and history’, whose songs contain, on average, 478 words. The graphs in fig. 2 give a more detailed insight in the distribution of songs within a genre. For the twelve most popular genres, the number of words per song is plotted against their frequency.²⁷ In general, all genres show a similar distribution: most songs contain between one hundred and three hundred words, with few songs boasting very long word counts. While the genres ‘religion’, ‘love and sex’, and the songs categorized with ‘none’ show a similar distribution, the word length within the genre ‘politics and history’ is more erratic. Compared to other genres, such as ‘amusement’, ‘cycle of life’, ‘groups’, and ‘children’, it shows a much wider spread in song lengths. It should be noted that in topic modelling the length of a text might influence its topic probabilities: because a long text contains more words, chances are it contains more topics than a short text.

One consequence of the piecemeal, project-based process through which this database has been constructed is a skewed representation of certain epochs. For example, while every Dutch monophonic song dated before 1600 and known to exist is included in the database, coverage of later periods is less complete. We also have to take into account that there is likely a genre bias to be uncovered in the compilation of songbooks – songs did not have similar chances of survival. Another drawback is that even when all known songs from a certain area are entered in the database, this probably is just a snippet of all songs that have been produced during that time. We can assume that a large fraction of the early modern song production is nowadays unknown to us, either because the documents no longer exist, or because they have not been recovered yet. Furthermore, the survival chances of a printed song are considered to be higher than that of a song in manuscript, and that of eighteenth-century song higher than that of a sixteenth-century song. In recent years, scholars have further explored to what extent the level of loss of historic works can be estimated. Leo Egghe and Goran Proot, for example, have proposed a probabilistic model, based on the frequency with which retrieved copies of historic works survive.²⁸ Mike Kestemont and Folgert Karsdorp have used an unseen species model from ecodiversity to estimate the survival rate of Middle Dutch chivalric epics.²⁹ Although this article will not contribute to the study of survival chances of early modern songs, it is important to note that the corpus I have used is undoubtedly biased and incomplete. Furthermore, these biases potentially change over time, and in that case also have implications for a comparative analysis of topics over time.³⁰

Topic modelling is a way of extrapolating backwards from a collection of documents to infer the discourses that might have generated them. It is assumed that each document in a collection of documents is constructed from a mix of some set of possible topics. A topic can be understood as a collection of words that have different probabilities of appearance in passages discussing the said topic. The model assigns high probabilities to words and sets of words that tend to co-occur in multiple contexts across the corpus. It must be noted that a topic is a probability distribution over the entire vocabulary, and that these distributions can vary hugely between topics. The ten most frequent words of one topic might together make up fifty percent of the topic, while the ten most frequent words of another topic only cover five percent of the topic. Although looking at the ten most frequent words of a topic is common in most topic modelling papers, it is in fact more insightful to look at the number of words of which the sum of probabilities equalizes a certain percentage.

Topic modelling is an unsupervised method, which means that the model makes ‘a wager that patterns in the data are sufficiently strong that different latent classes of observations will make themselves “visible” ’.³¹ This means that the algorithm is not informed about the kind of texts that are imported and what themes it has to look for. The algorithm infers information about individual word meanings based on their repeated appearance in similar contextual situations.

For the purposes of this study, topic modelling was performed using the Python package gensim, which contains a wrapper for mallet, a Java topic modelling toolkit. I used the statistical technique Latent Dirichlet Allocation (lda), which is implemented in mallet as well as in other software packages.³² lda is the most commonly used topic modelling method. ‘Latent’ refers to the concept of latent distributions in the model, and ‘Dirichlet’ is the name of the most commonly used prior distribution, named after the German mathematician Johann Dirichlet. It consists of a nested multilevel structure, in which the three levels – word, document, and corpus – are distinguished.³³

Since topic modelling is an unsupervised method for automatically extracting topics from unlabelled data, it can give no guarantees on the reliability and interpretability of the output. A method to measure the coherence of the topics is therefore highly desirable. A coherent topic is a topic that is semantically interpretable: if a topic consists of the words Jezus, kruis (cross), lijden (suffering), and Golgotha, we are clearly dealing with the crucifixion of Christ. However, a topic can also be a result of either the model’s design settings, or of artefacts in the corpus, such as selection criteria leading to an imbalance in represented genres, resulting in topics in which this imbalance is reproduced. This can (but does not necessarily have to) result in a topic that is not semantically interpretable. An example would be a topic consisting of the words kribbe (manger), prins (prince), Holland, geus (beggar), Venus, and minne (love). A coherence measure indicates whether the formed topics are indeed semantically interpretable.

In a 2015 study, Michael Röder, Andreas Both, and Alexander Hinneburg evaluated the performance of several coherence measures. They concluded that the best performing coherence measure was a rather unexplored one, which they labeled C_v.³⁴ C_v is based on four parts: i) segmentation of the data into word pairs; ii) calculation of word or word-pair probabilities; iii) calculation of a confirmation measure that quantifies how strongly a word set supports another word set; and iv) aggregation of individual confirmation measures into an overall coherence score.³⁵ The coherence score is a value between 0 and 1. In general, a higher coherence score indicates a better quality of topic model.³⁶ To determine the settings for some parameters that have to be set in building a topic model, I used C_v as coherence score.³⁷

In order to build a topic model with gensim, a few preliminary steps have to be taken. An important first step is to reduce spelling variation. The early modern period lacked standardised spelling and grammar, which meant that (and was proliferated by the fact that) Dutch authors and printers attached little value to consistency in spelling, resulting in the coexistence of multiple variants between, and sometimes within, texts. Because topic modelling is based on the position of a unique word in a text (and thus each unique spelling variant) and its distribution through a corpus, spelling variation disturbs the results of the process. Reducing the variation in spelling is therefore a crucial step that needs to be taken. For the English language, several tools have been built over the last years to convert historical spelling variants into their modern equivalent. For the Dutch language, however, such options are limited, and the tools that are available come with both advantages and disadvantages. For the purpose of this study, I used the Dutch setup of the variant Detector (vard), a semi-automatic tool introduced by Alistair Baron and Paul Rayson in 2005.³⁸ The tool uses two lists (a normalized word list and a variant list) to suggest or replace variant words with their normalized counterparts. The normalization suggestions are given based on a combination of four different methods: i) known variant replacements; ii) character edit distance; iii) letter rules; and iv) phonetic distance. The Dutch version of vard2 was built by Ivan Kisjes and Tessa Wijckmans, and was the result of a collaboration between the create-project of the University of Amsterdam and the Dutch digital research platform Nederlab. As a training set, they used the first two books of the 1657 edition of the Dutch translation of the Bible.³⁹

vard2 will find numerous potential normalizations for a given variant. Each of these suggestions is given a confidence score, in the form of a percentage expressing how certain the tool is that this is the correct normalized word. The suggested normalization with the highest confidence score is used to replace each variant. Consider the opening lines of a song from 1598: ‘Reyn maecht eerbaer/aenhoort myn claeghen snel’ (‘Pure honourable virgin, hear quickly my complaint’). vard2 suggests normalizations for every word, except the word snel (quick). The word reyn (pure) gets as most popular suggestions the words rein (98.47 percent) and reen (0.39 percent). The percentages show that in this case, rein (clean) is clearly the correct normalization suggestion. For the word maecht (virgin), however, the tool offers several potential candidates. The words maagd (virgin, 95.59 percent), macht (power, 42.71 percent), and maakt (makes, 38.52 percent) are suggested as normalizations, although the high percentage corresponding to maagd shows that the tool is still confident that this should be the correct normalized word.

However, if the system’s normalization methods struggle with a particular variant, the highest confidence score may be relatively low – in these cases a threshold is required, which is the minimum confidence score needed for normalization to take place. If the threshold is not met by the top normalization suggestion, the word will not be normalized, and will be left as a variant. The word claeghen (to complain) from the earlier-cited song gets as suggested variants clean (0.96 percent) and clan (0.93 percent). Clearly, klagen is not included in the normalized word list. If the threshold is higher than one percent, the word ‘claeghen’ will remain unnormalized. A higher threshold will increase precision but reduce recall. After using a single text and checking the number of words that were normalized at different thresholds, I decided to set the threshold at fifty percent. Before normalization, the corpus consisted of 217,906 distinct words; after normalization with vard2 the number of unique words in the corpus was reduced by 23,857, resulting in a corpus with 194,049 distinct words.⁴⁰

The next pre-processing steps included tokenization – meaning that each text was converted to a list with separate words, or tokens – and punctuation removal. Subsequently, stop words, which are the very common words (in this case mostly function words such as personal pronouns, propositions, and conjunctions), were removed from the corpus. Using the Python package nltk, a list of one hundred and fifty most frequent words was made. Content words were manually removed from this list and therefore kept in the corpus.⁴¹ Furthermore, a threshold was set for words that appear too infrequently in the corpus (no_below). Calculating the coherence score of topic models with different settings for no_below, this threshold was set at 2, which means that words that appear in only one document are ignored.⁴² The third step was deciding how many topics the model would be creating. Here the coherence score was used again. I built ten topic models with different values for the number of topics, in the range 10:100. The coherence score of each model was calculated. It turned out that the model with fifty topics obtained the highest coherence score, and therefore I decided to build the final topic model using this setting.⁴³

The output of a topic model consists of two dataframes (tables): one with the keys of a topic model, the other one with the composition. The dataframe ‘keys’ contains fifty rows, one for each topic, and two columns: one for the topic id, the other for the words that form a specific topic. The dataframe ‘composition’ contains as many rows as there are documents in the corpus. The columns are numbered from 0 to 49, corresponding with the id of the topics from ‘keys’. The cells from each row contain a value between 0 and 1. The higher a value, the more dominant that topic is in a document. Note that the sum of each row will always be 1, but that the distribution always differs.

After building the final topic model, another tool was used to measure the quality of the model, and to get more insight into the topics and how they are related to each other. I used the Python package pyldavis, which is a library for interactive topic model visualization. The package extracts information from a fitted lda topic model to inform an interactive web-based visualization. After a few steps of pre-processing, the output is an Intertopic Distance Map. Fig. 3 gives an example of the data visualized, in this case topic 27 (‘world & money’). Each bubble on the left-hand side of the plot represents a topic.⁴⁴ The larger the bubble, the more prevalent the topic is within the corpus. A good topic model should have bubbles scattered throughout the chart, rather than clustered in a single quadrant. A model with too many topics will typically have many overlapping small-sized bubbles clustered in one region of the chart. When hovering over the topics, the thirty most prevalent words of that topic are visualized on the right-hand side. Since a topic model only produces bags of words as output, users have to manually interpret the words that make up a topic. To do so, I looked at the number of words that together made up fifteen percent of a given topic. I built a word cloud from every topic, visualizing the upper fifteen percent of words and their weights within a topic. The top thirty words and their weights within every topic are included in an online appendix.⁴⁵ The topics and the subjects I assigned to them can be found in tab. 2.

The appendix also includes a column with the sum of the weights of the top thirty words. The lower this value, the ‘flatter’ the distribution of words within a topic, suggesting that the words in these topics are not typical for that specific topic, but occur in many other topics as well. The clearest example of such a word is ‘God’: this word has the highest weight in ten topics. Still, its probability can be fairly low (0.00992 in topic 7, 0.0131 in topic 28). An example of a topic with a high sum of weights of the first thirty words is topic 29 (0.2493), to which I assigned the label ‘drinking’, due to words such as wijn (wine), drinken (to drink), vrolijk (merry), drinkt (drinks), and bier (beer). Another clear example is topic 46 (0.2620), to which ‘life & sin’ was assigned, because of the presence of words such as wereld (world), God, leven (life), kwaad (evil), and sterven (to die). Topic 5 was labelled as ‘Christmas’, due to words as kindeken (baby), stal (stable), Bethlehem, and Maria. I assigned ‘love & sadness’ to topic 39, because of the presence of positive words such as lief (sweetheart), liefde (love), min (love), and schone (beautiful) on the one hand, and negative words such as ach (oh!), pijn (hurt), smart (grief), and verdriet (sorrow) on the other. This topic differs from another topic on love (topic 38) in that almost all the words in this topic are positive: hart (heart), vreugd (joy), leven (life), min (love), and vermaak (entertainment). I assigned ‘love & happiness’ to this topic.

I labelled topic 32 ‘nation & country’: the words prins (prince), stad (city), land (country), graaf (count), and Holland leave no doubt that these are words from political songs. However, the sum of weights of the first thirty words in this topic is rather low: 0.1342. The same goes for topic 14. While the words all refer to a topic that I labelled ‘myth & beauty’ (including Venus, schoonheid (beauty), gezicht (face), glans (shine), and goden (gods)), the sum of weights is only 0.1429. This suggests that these topics have a flat distribution. In topic 47, words from pastoral or bucolic songs are clustered. Many words refer to the shepherd’s life, such as vee (cattle), herder (shepherd), bos (forest), and schaepjes (little sheep). I assigned the label ‘bucolic’ to this topic. Topic 28 was labelled ‘prayer & preaching’, because apart from the obvious religious words, there are words in this topic that have an oral connotation – hoort (hears), horen (to hear), zeid (said), verstaat (understands), vermaan (admonish), and broeders (brothers). It turns out that many topics have something to do with religion. This is not surprising, since we already know that almost forty percent of the corpus are songs from the category ‘religion’. In some cases, it is quite easy to define a label for a topic (such as ‘religion & happiness’, ‘religion & sin’), but in other cases the differences are more difficult to distinguish.

To examine which topics dominate the corpus, I selected for each song the topics that covered at least ten percent of the song, meaning that a topic has a minimal weight of 0.1 in a song. Subsequently, I summed these weights per topic (see tab. 2). The topics are ranked by the sum of their weights. Topic 9 is by far the most popular within the corpus. The appendix shows this topic has a rather ‘flat’ probability distribution: the five most frequent words in this topic (god, wt, leun, gods, nv) all have a weight of between 0.01 and 0.02. The oddly spelled words such as wt, leun, nv, si, bouen, and ouer suggest that topic 9 is rather a result of artefacts in the corpus: the tool used for spelling normalization does not work well on texts written before 1637. Fig. 4 confirms this hypothesis: topic 9 peaks between 1550 and 1600, but descends soon after.

The rest of the top ten of most dominant topics is made up by topics that relate either to religion or to love, except topic 19, in which words are not clustered on semantic similarity, but on linguistic characteristics. I labelled this topic as ‘verbs’, because it consists of words such as leven (to live), zullen (shall), geven (to give), and komen (to come). The relatively low sum of weights of the first thirty words (0.1573) indicates that these words are not very distinctive for this particular topic, but occur in many other topics. If we take a look at the bottom of the chart, the opposite is the case. Topic 41 consists of German words (das (the), mein (mine), mir (to me), ein (a/an), mit (with)) and has a sum of weights that is above average (0.2697). The same goes for topic 43, which consists of French words and has a sum of weights of 0.3463, as well as for topic 20, consisting of interjections that often occur in children’s or folklore songs, with a sum of weights of 0.5108. These high values indicate that these words are rather distinctive for these topics.

The most dominant topics in the corpus, then, are ‘religion plus something’. This is not a surprising result, because the distribution of categories already indicated many religious topics. Still, the topic model has uncovered many differences within the wide-ranging genres of songs. Although many more observations can be made with regard to the fifty most distinctive topics, the next two sections will explore how the dominance of song topics fluctuates in the context of the two cases I introduced earlier: Petrarchan songs and patriotic songs. As we have seen, these are topics that are popular within the corpus, and based on qualitative research we would expect to see some changes over time regarding their popularity. A quantitative approach can be used to detect this.

Based on the existing literature, song topics related to the Petrarchistic tradition should be peaking in the beginning of the seventeenth century. These could be topics on the tragic aspects of love, but also topics which can be related to the Renaissance and classical tradition, for example the bucolic genre. Two dominant topics in the second-largest overarching category – love – are contradictory: topic 39, ‘love & sadness’, and topic 38, ‘love & happiness’. Fig. 5 shows the evolution of these two topics over time, which reveals an interesting pattern: for instance, the topic ‘love & sadness’ peaks about fifty years before the topic ‘love & happiness’.⁴⁶ Initially, topic 39 is dominant, but by 1700, topic 38 (‘love & happiness’) picks up and rapidly becomes the dominant of the two love topics. The topic ‘love & sadness’ does increase slightly in the eighteenth century, but never comes close to ‘love & happiness’. ‘Love & sadness’ can be seen as a typical Petrarchan topic: the words ach (woe!) and pijn (pain) in this topic speak for themselves. Fig. 5 demonstrates that the popularity of these songs peaked in the early seventeenth century, as expected. It also suggests that towards the end of the seventeenth century, songs with negative emotions were replaced by songs with a rosier view on love.

Fig. 6 shows additional love topics besides the two discussed above, demonstrating that another Petrarchan topic (2, ‘love & tragedy’) had in fact peaked in the first quarter of the seventeenth century, before being replaced by topic 39. When we take a closer look at the words that form the latter two topics, it becomes clear that they exhibit many similarities: both contain the words liefste (dearest), liefde (love), herte (heart), minnen (to love), pijn (pain), and verdriet (sorrow). However, both topics also contain distinctly different words: topic 2 includes words such as troost (consolation), lijden (suffering), genezen (to heal), and scheiden (to separate), while topic 39 consists of words such as smert (grief), waarom (why), droeve (sad), and sterven (to die). Semantically the topics are related, but they include different words. Taken together, the two topics suggest the dominance of tragic love over all other love tropes in songs published between 1575 and 1685.

Topic 24 (‘rejection’) peaked simultaneously with ‘love & sadness’, between 1650 and 1675. Other love variants, such as ‘seduction’ and ‘physical love’, are less prevalent, but still peak at the same time as topics 39 and 24. These topics may also be connected to the heyday of Petrarchism. A famous Dutch imitator of Petrarch was Pieter Corneliszoon Hooft, whose poems were largely based on the Petrarchan exemplar, in particular the conventions of amorous lyric. The case of Hooft demonstrates that not only poems but also songs were written with the Petrarchan tradition in mind. Hooft’s songs ended up in songbooks, accompanied by melodies, and reached a large audience. From 1607, song poets often referred when assigning their melodies to songs from Hooft’s play Granida (1605).⁴⁷ These observations are reflected in fig. 6, which shows that various love topics, which in their own way all relate to Petrarchan love poetry, dominated between 1575 and 1650.

Other topics related to the humanistic tradition are ‘bucolic’, ‘myth & beauty’, and ‘nature’, which are visualized in fig. 7. ‘Bucolic’ refers to the pastoral, a genre originating in the classical tradition and especially in the works of Theocritus and Virgil. The use of mythological figures such as Venus and Phoebus in topic 14 also indicates the imitation of classical genres. There is a clear decline in these topics after 1660. ‘Love & sadness’, ‘love & tragedy’, and ‘rejection’ are overtaken by ‘love & happiness’ (fig. 6). The Petrarchan tradition with negative emotions on love seems to have moved to the background, making room for a less dramatic, more light-hearted interpretation of love. An explanation could be that the moralistic attitude described by Keersmaekers, which can be linked to the culture of piety that was blooming in the seventeenth century, was beginning to diminish. Qualitative research could provide more insight into whether (and if so, why) the depiction of the Petrarchan tradition in songs declined in the course of the eighteenth century.

Scholarship has suggested that topics dealing with nation and country peaked during the Dutch Revolt, although they remained dominant after the 1648 Peace of Münster. I have selected three topics related to politics to test this hypothesis. Fig. 8 shows that topic 32 (‘nation and country’) appeared around 1575 and steadily rose thereafter. The topic remained dominant for at least fifty-five years, a timespan largely coinciding with the Dutch Revolt. Dominant words in this topic are prins (prince), land (country), graaf (count), spaensche (Spanish), and paus (pope), confirming that these topics are indeed covered by the beggar songs, which were extremely popular during the Dutch Revolt. It seems likely that the many reprints of the Nieu Geusen liedtboecxken (published between 1577 and 1687) are responsible for the dominance of this topic. The song that is most emblematic for topic 32 (because of its high weight) is titled ‘Wel op wel op Spangiaerden’, an anti-Spanish song from the Nieu Geusen liedtboecxken.⁴⁸ However, more songbooks with patriotic songs existed in the Dutch Republic.⁴⁹ Although some of the songs in these songbooks were also included in the beggar songbooks, others that would have fitted the criteria were not. Topic 32 shows a drastic fall around 1640, which coincides with the Peace of Münster in 1648. The topic is no longer dominant in subsequent decades, bar a little revival between 1675 and 1725. Since all graphs use a rolling average over twenty years, this revival could be the result of the last reprint of the Nieu Geusen liedtboecxken in 1687. All in all, fig. 8 suggests that the dominance shown by this topic during the Dutch Revolt did not return, merely that there is perhaps a renewed popularity of patriotic songs during the War of the Spanish Succession (1701-1714), though the line’s erratic course demands that caution be shown in making such an interpretation.

Around 1625, when topic 32 began to decrease, the two others gained in popularity, of which ‘God and enemy’ became even more dominant than topic 32. An important contrast between topics 7 and 49 on the one hand, and topic 32 on the other, is the presence of religious words in topics 7 and 49. The only religious word in topic 32 is God, with a considerably lower weight, given that words such as prins (prince), land (country), stad (city), and vijand (enemy) are more prevalent. Topics 7 and 49 contain more overlap with other religious topics, with the presence of such words as God, boze (angry), Gods (God’s), and handen (hands) in topic 7, and God, zijne (His), naam (name), heren (the Lord’s), and hemel (heaven) in topic 49. This suggests these are religious songs with a strong focus on the aspects of war, peace, enemy, and the people. The song with the highest weight contribution to topic 49 turns out to be a psalm in a translation by Petrus Datheen; indeed, many other songs with a high weight contribution to topic 49 are also psalms. This is not surprising, as many psalms use a vocabulary that expresses a battle against a real or imaginary enemy. The song with the highest weight contribution to topic 7 is from the songbook Stichtelijcke Gesangen (1661) and relates the story of the biblical figure Daniel, who was thrown into a lion’s den:

Although the first verse relates Daniel’s well-known fate, the second contains a reference to a war the singer was involved in. This not only ties in with the aforementioned increase in attention paid by orthodox ministers to patriotic ideas, but also with a broader trend in orthodox circles: during the Further Reformation (a Pietistic movement that began in the second half of the seventeenth century), songs were increasingly considered to be a literary genre through which the dissemination of Pietistic ideas could take place. Proponents of the Further Reformation published their own catechisms, sermon collections, and songbooks, which caused an exponential rise in the number of Reformed songbooks after 1650.⁵¹

Another explanation for these trends could be that these topics, in addition to their political connotation, also have a military meaning. Furthermore, the topics might also be (partly) formed of songs in which earlier military events are historicized. In summary, the topics related to the fatherland not only show how these songs prevailed during political conflict, but the patriotic ideas they embodied also became increasingly embedded in the seventeenth-century Pietistic tradition.

This article has shown that computational methods for text analysis offer a meaningful way to analyse a large digital corpus. The benefits of this study are, therefore, multiple: first and foremost, topic modelling has provided a possible solution to the issues addressed in the introduction – the lack of metadata on genres in the Dutch Song Database, and the impossibility of searching for genre patterns in this collection. Although the dominance of religious and love topics was to be expected, given the existing genre classification of the database, this study shows that subcategories can be defined within a genre. For example, the built topic model was able to distinguish between several aspects of love, different focuses within the topic ‘religion’, and differences in songs with political connotations. These results contribute to a better understanding of the richness of the Dutch Song Database. A recommended next step would be to supplement the missing metadata on genres in the database with the results from this study, as this would facilitate its use by other researchers.

Secondly, by exploring the relationship between topical fluctuations and cultural-historical changes in the Dutch Republic, I have shown how quantitatively obtained results can be related to existing scholarship on early modern Dutch song culture. This study confirms that Petrarchan ideas were indeed transmitted through song, a phenomenon that began in the early seventeenth century. Furthermore, the results suggest that the Petrarchan tradition in songs started to decline in the course of the eighteenth century. In addition, this study shows that songs with a patriotic identity became more popular during political crises, and that songs belonging to the religious genre were also used to spread patriotic ideas, a trend coinciding with the heyday of the Further Reformation.

More generally, I have shown how, with the methodology used in this study, diachronic developments in textual corpora can be explored. This particular study shows how song lyrics interfere with and react to socio-cultural and political developments, thus demonstrating that songs are a suitable medium through which the historian may gauge the social climate. The question rises, however, whether songs are precursors to or merely the result of socio-cultural trends, something that deserves further research. Furthermore, the proposed method should be improved, so that it may be applied to other corpora, both early modern and modern, both in Dutch and other languages.

A few remarks should be made regarding the methods used and the results obtained. In the first place, this study shows the complexity of using historical textual data on a large scale. Although retaining spelling variation might be necessary for other studies or disciplines, in this case it was important that spelling be normalized across the corpus. The method used to reduce spelling variation in this study is not flawless. For example, because the training set of the tool was built from seventeenth-century material, contemporary words were not normalized by the tool. This resulted in topics that were not semantically interpretable; rather, a group of word clusters formed that were based purely on the fact that their words were not normalized. Clear examples are topics 9 and 45, which contain words such as dij, yn, soe, uwz, jon, and fen. This shows once again that more research into techniques to deal with historical variation in Dutch spelling is crucial.

It must also be noted that topic modelling is an oft-debated computational method, not only because it is based on probabilities and the mathematical idea behind the method seems a mystery to many users, but also because it is clear that more work is required on measures to evaluate the results. In this study, evaluation was done by using coherence measures and visualizations, but it is important to stress that these evaluation methods are also not flawless. Another point where intervention of the researcher may have led to biased results was the assigning of a description to topics. This was done by looking at the word cloud of a topic’s fifteen percent most dominant words, but it is entirely possible that another researcher would have assigned different subjects to the topics.

In general, it should be noted that when using topic modelling, songs originating from a certain context, and sung by multiple groups of peoples for various reasons, are simply reduced to groups of words. It is always necessary, therefore, to interpret and evaluate the results of a topic model in detail, as well as put them in the cultural, sociological, and historical context in which the texts were produced. I have tried to do this to a certain extent, but further research can undoubtedly improve my findings. The same goes for other aspects of songs and song cultures that were not taken into account in this study, but that can certainly affect a song’s popularity, such as its musical style and communal singing practices.

Scholars have published many qualitative studies on well-known and lesser-known songs, songbooks, and song writers from the Dutch Republic. The aim of this study, in contrast, has been to offer a quantitative perspective, and to demonstrate how the combination of both approaches can offer new perspectives to the field of early modern literary studies. My quantitative research has generated new hypotheses about the Dutch song culture that merit investigation, both by quantitative and qualitative literary researchers and cultural historians. I believe that zooming out, while inevitably rendering some details invisible, allows us to see new patterns and generate new hypotheses that can improve our understanding of not only early modern song culture, but Dutch culture at large.