Between the Middle Ages and the nineteenth century, many middle-class Europeans kept a handwritten chronicle, in which they reported on events they considered relevant. Discussed topics varied from records of price 昀氀uctuations to local politics, and from weather reports to remarkable gossip. What we do not know yet, is to what extent times of con昀氀ict and crises in昀氀uenced the way in which people dealt with information. We have applied methods from information theory - dynamics in word usage and measures of relative entropy such asnovelty and resonance - to a corpus of early modern chronicles from the Low Countries (1500-1820) to provide more insight in the way early modern people were coping with information during impactful events. We detect three peaks in the novelty signal, which coincide with times of political uncertainty in the Northern and Southern Netherlands. Topic distributions provided by Top2Vec show that during these times, chroniclers tend to write more and more extensively about an increased variation of topics.
Between the Middle Ages and the nineteenth century, many middle-class men (and a handful
of women) in Europe kept a handwritten chronicle, in which they reported on current events
in their communities, and on what they considered interesting or relevant. These texts were
ordered chronologically, providing both the date and a report of a certain event (seFeigure 1).
Chronicles were rarely printed in the lifetime of their authors, but, despite their scribal form,
they could still circulate in the localities, be read and continued by other authors, in昀氀uencing
future generations [
The described topics in chronicles were various. Chroniclers alternated between
descriptions of political developments and records of price 昀氀uctuations of grain, butter, and milk,
weather reports, mentions of their relatives’ birthdays, gossip, religious developments, and
reports on unusual, strange, or marvelous events both nearby and further away. Although
chronicles did change over time, the type of information chroniclers selected to be included
remained fairly stable, which enables us to study the genre across centuries. One subcategory
of chronicle texts emerged out of political crises, wars, and civil con昀氀icts. During such times,
many authors started to record public events that were upsetting their lives and to teach their
readers the lessons to be learned from them [
What we do not know, is how times of con昀氀ict and crises in昀氀uenced the chronicler’s way of reporting and dealing with information. Did chroniclers collectively write about the same topic during such a crisis, or were they more recipients of broader information? This paper will demonstrate how chroniclers were coping with information during impactful events. We aim to provide more insight into the way early modern people dealt with crises. We apply methods from information theory to a corpus of early modern chronicles from the Low Countries to 椀昀nd an answer to the question formulated above.
Related work will be discussed in Section 2. The used corpus will be introduced in Section 3, while Section 4 contains a description of the methods used in this study. In Section 5, the obtained results will be discussed, and in Section 6, we draw some conclusions from the results and do some suggestions about the direction future research should take.
Chronicles can be considered as a collection of chronological events. However, few historians
have been studying ‘the event’ as a theoretical category [16, p. 198] – even though the
antievenementalism of social historian Fernand Braudel, who considered the history of events as
the mere froth on the waves of history, was largely replaced by a return to writing about events
in the 1970s. Anthropologist Marshall Sahlins has stated that ‘[e]vents can be distinguished
from uneventful happenings only to the extent that they violate the expectations generated by
cultural structures. The recognition of the event as the event, therefore, presupposes structure’
[16, p. 199][15]. The same distinction is made by [
From a linguistic perspective, the question of whether an incident is important enough to be
an event is not relevant: every state, change, or happening is considered an event. In linguistics,
the term event is related to the concept ofeventuality, which was introduced by the linguist
Emmon Bach in 1986 and comprised states, processes, and events 2[]. Many linguists indeed
understand the term eventuality in the broadest sense, comprising events, processes, states,
happenings, changes, episodes, etc., as is for example the case in a study on event detection
and classi昀椀cation for historical texts [
The di昀erence between the historical and linguistic perspectives on events is that linguists focus on the linguistic elements that are used, while historians focus on the result of an event. When linguists use subcategories to distinguish between events, it is based on these linguistic elements. Neither the historical perspective nor the linguistic perspective on events is completely suitable for the study of events in early modern chronicles. In these texts, both abstract and emphatic events are included. First and foremost, chroniclers report on happenings in their surroundings that they consider being noticeable ruptures of their daily routine. This can be the threat of war, the visit of a foreign king, a national con昀氀ict, or unusual weather phenomena. This could be considered an event in the emphatic sense. At the same time, however, chroniclers also include reports on daily or weekly routines such as checking the wind direction and temperature, summarizing the Sunday sermon, reporting on the prices on the market, or referring to weekly board meetings they attend. These are the abstract kind of events. In the context of this study, we consider an event to be every description that is linked to a speci昀椀c date. For more details on methods see Section 4 and Appendix A.
The methods used in this study are based on dynamics in word usage and measures of relative
entropy. We know from previous work that word usage in newspapers is sensitive to the
dynamics of socio-cultural events [
Previous studies have shown that entropy measures can be used to detect fundamental
conceptual di昀erences between distinct periods [ 7], opposite political movements [
The data set applied in this study is a subset of a corpus that was collected and digitized in the context of the research project ‘Chronicling Novelty. New knowledge in the Netherlands, 15001850’.1 The full corpus consists of about 320 early modern chronicles that are written in the Dutch language between 1500 and 1850. They are chronologically organized, cover events that happened in the lifetime of the author, and focus on local events more than national, individual, or familial. About 130 of these chronicles had been published before as a contribution to a journal, on the initiative of an archive, or in the private domain, and were digitized by the Digital Library for Dutch Literature (DBNL). The other chronicles are kept in libraries and archives throughout the Netherlands and Belgium.
Every manuscript page was scanned and transcribed with both the Handwritten Text
Recognition tool Transkribus [
Our data set consists of 191 chronicles that were fully transcribed and annotated. However,
the date tag, which plays a pivotal role in this study, repeatedly was subject to many bugs and
crashes. The tag therefore needed a manual inspection. We use all 191 annotated chronicles for
training models (corpus annotated), but only the 96 chronicles in which thedate label was
manually checked, were used for analysis (corpus corrected). See Table 1 for more statistics
on the used corpus.
1On http://www.chroniclingnovelty.com/kronieken,/ an overview of the corpus can be found.
4. Pipeline
We developed a 昀椀ve-step research pipeline, additional details can be found in appendix A2:
1. Chunk chronicles into primitives. Since chronicles are chronologically structured
and mentions of a date are labeled as such, we make a cut before everydate label, to
chunk the texts into smaller fragments that can be connected to a date label. This was
done for bothcorpus annotated and corpus corrected. We call the resulting text
chunks primitives. We furthermore made a subset of primitives corrected daily,
only containing primitives with a fully speci昀椀ed date tag. Table 2 contains statistics
of the distinctive data sets. Theprimitives annotated are used in step 2, while the
primitives corrected daily are used in the other steps.
2. Primitive representations. We use a Top2Vec model to create both document
representations of the primitives annotated, and topics [1]. The model provides two relevant
outputs, which are the doc2vec embeddings of the documents [
We expect peaks in the novelty signal to be indications of an event. 2Please refer to the git repository for the full code:https://github.com/centre-for-humanities-computing/dutch-c hronicles.
3Given this, it is striking that no peak appears in 1672, the DutchRampjaar (Disaster Year) in which the people were described as redeloos (irrational), the government as radeloos (distraught), and the country asreddeloos (beyond salvation). However, this absence is mainly due to a lack of data in the full corpus from that period. words wolkachtig motbuitjes verhelderd betrekt wolkens buiachtig verdund verdonkert betrokken overleede hat do overleeden edl weeduwe mevrou weed ju昀rou haar jaaren niwe oostindise dri par la dont dans dernier nouveau un alors le sur avec du francais on il etre une consequence les weinig duur daardoor thans hooi mogelijk steeds aanzien vooral prijzen aardappels oorzaak spaengiarden ducdalbe vuel ducdalf deestyt brusel vuele prinsche dagelycx scuyten gescut plechtigheyd geluyd beyaerd triumph luyster feest autoriteyten bywezen magten musicq Several categories can be discerned from qualitative inspection of the topics in the reduced Top2Vec model. There are natural, cultural, social, economic, and political topics – and some topics fall into multiple categories. A sixth category contains topics of words that are not clustered on semantic similarity, but on linguistic characteristics. InTable 3, we included an example of every category. Topic 0 (the most dominant topic within the corpus) clearly belongs to the category of natural topics, containing words related to the weather. Topic 1 falls into the category of social topics, with words such as ‘overleeden’p(assed away), ‘weeduwe’ (widow) and ‘ju昀rou’ ( miss). Topic 2 is not really a topic, because its words are clustered on linguistic characteristics: they are all French. In topic 39, words related to prices and products, such as ‘duur’ (expensive), ‘hooi’ (hay), ‘aardappels’ (potatoes), and ‘prijzen’ (prices), are clustered, belonging to the economic category. The words ‘spaengiarden’S(paniards), ‘ducdalbe’ (Duke of Alba), ‘prinsche’ (prince), and ‘gescut’ (artillery) indicate that topic 42 can be considered a political one. Topic 48 contains words related to festivities (‘plechtigheydc’e(remony), ‘beyaerd’ (carillon), ‘feest’ (party), ‘musicq’ (music) and thus belongs to the cultural category.
Topical dynamics were represented by plotting the mean cosine similarity of all primitives in one year towards a certain topic. Some of the topics show a clear trend over time which can easily be linked to a social, cultural, or political situation at that time. Other topics demonstrate a repetitive pattern, remaining stable on the long term.
The prototype primitives, of which the frequency is visualized in dark green inFigure 2, were used to compute the novelty signal. Three peaks can be observed in the novelty signal in Figure 3, which remain solid when we adjust the window size . Peaks are visible around 1568, 1662, and 1789. These 昀氀uctuations in the signal indicate that a document vector is surprising compared to its preceding vectors. The valley the novelty signal shows a昀琀er a peak means that the following documents are less surprising because their vectors are more similar to the previous ones. The slow ascent of the peaks in 1568 and 1789 point to a long period of steadily increasing surprise. The steep descent that follows, indicates a sudden decrease in content novelty of the documents.
The three peaks mark the earlier mentioned key points in the history of the Low Countries: the start of the Dutch Revolt (1568), the prelude to the War of Devolution (1662), and the end of the patriot movement with the return of the Stadtholder (1789). Furthermore, they approximately coincide with a peaking primitive frequency. This indicates that when more chroniclers report (or chroniclers report more) in their chronicle, the content of their reports changes. A peak in the novelty signal can mean several things, for example,(1) a few topics become more dominant than others,(2) earlier dominant topics become less dominant, or(3) more diverse topics appear. The yearly mean cosine similarity per topic that was obtained with Top2Vec is used to get insight in the topics that are dominating at the time of the novelty peaks. Although one might expect that, during crises, chronicles would become more focused on a small number of topics directly related to the crisis, it turns out that the topic distribution is 昀氀atter during novelty peaks than in other years: even the topics with the highest mean cosine similarities are still close to or below 0. It indicates a large variety in described topics during such years. The top words of the three most dominant topics during novelty peaks are included in Table 4.
It must be said that document length is an important driver for the variation between high and low novelty. Longer documents are more novel than shorter documents, because long documents contain more varied information and therefore have average similarity to many topics, while short documents, containing less information, have high similarity with one topic, but uniformly low similarity with the rest.
A positive association between novelty and resonance would indicate an innovation bias, meaning that novelty introduced in the past leaves traces in the future. In our results, the resonance signal remains 昀氀at over time, which suggests that future information does not conform to the introduced novelty. It means that a newly introduced event in the corpus of chronicles does not have an e昀ect on the content that is described a昀琀erward. In other words: big events do not impact the writing style and habits of chroniclers in the long term.
Di昀erent interpretations are possible regarding the obtained results. Increasing diversity in described topics during crises suggests that during times of high uncertainty, people are more open to new information. Alternatively, it can be an indication of an expanding mediascape, or a more thorough consumption of new media. Furthermore, it may be indicative of how early modern people understood crises in a di昀erent way than we expect. A crisis is not only about soldiers, sieges, and deaths (topics 4 and 5 in 1662), but it is also relevant what kind of food is available (topic 10 in 1568), how the weather might in昀氀uence this (topic 0 in 1789), what the city government decides (topic 9 in 1568), how the situation in a foreign city evolves, and to not forget religious duties (topic 49 in 1789). This variety in topics could also show that not every chronicler was equally a昀ected by these historical events. Still, they felt the need to start or intensify their recordings on events happening in their lives. An overload of information asks for a su昀케cient approach. A昀琀er being exposed to an information explosion, people tend to select what is relevant to them, discarding the other topics. This is what the resonance signal shows: there is a short period of information overload, but soon, the variety of described topics decreases again.
We have presented a method to detect events in early modern Dutch chronicles, which has provided insight into the way chroniclers cope with information during impactful events, and how these happenings in昀氀uence their way of writing. Our main conclusion is that early modern chronicles tend to write more and more extensively during times of political uncertainty. However, the topics they describe during such times also get more varied. This is shown by a peak in the novelty signal, and a 昀氀atter topic distribution. Furthermore, such an increase in event density and a change in the novelty signal does not in昀氀uence future reported events. Soon, things get back to normal, and the (writing) life of the chronicler continues as it did before.
The representativeness of the corpus used for analysis is not unproblematic and is something that could be improved in future research. We have used about one-third of the full corpus of chronicles, mainly due to the fact that only this part was digitized and annotated at the time. Besides, the focus on only ‘daily events’ introduced a bias in the results. The date with the highest frequency of primitives (67), most of them were from one author, describing how the Stadtholder Willem V visited his hometown Purmerend. Other frequently reported dates pointed us as well to events happening on one day, rather than to events spread over a longer period of time. An exploration of the frequently mentioned dates would gain value when ‘monthly dates’ were also included.
In this study, we have used the corpus as a whole in computing novelty and resonance signals, showing that the genre remains stable over time, despite of several impactful events. Future work will focus on the novelty and resonance signals of individual authors, in order to see whether the elevated topic diversity during crises can also be observed here. This should furthermore provide more insight in their personal writing style, and whether certain events they experience have a lasting in昀氀uence on their way of chronicling. Other future research should investigate whether the absence of a positive relationship between novelty and resonance is distinctive for the genre of chronicles, or more general something that can be observed in early modern texts. Early modern pamphlets or newspapers – speci昀椀cally written for an audience – would be an interesting showcase to explore their di昀erence from chronicles – since the latter are in the end more o昀琀en a case of a private matter.
This research was executed at the Center for Humanities Computing Aarhus. The research stay of Alie Lassche was funded by the Dr. Catharine van Tussenbroek Fonds, and the Stichting de Fundatie van de Vrijvrouwe van Renswoude te ’s-Gravenhage. Kristo昀er Nielbo was supported by the Nordic e-Infrastructure Collaboration (NeIC) and the Danish e-Infrastructure Cooperation (DeIC) with grant DeiC-AU1-L-000001.
The corpus does not come reliably segmented into sentences or documents. We use provided date tags for our document segmentation. In general, we consider a document to be the text beginning with a date tag<date A>, which in turn serves to date the document. A document can span multiple lines and pages, and ends with a date tag<date B>, which indicates the beginning of the next document. However, we do not segment within lines – a document always contains an entire line of text. In doing so, we attempt to address the cases, when an event’s dating does not appear at the beginning of the entry, but at some later point (e.g. in the middle of a line). As a result of this rule, lines with multiple date tags are considered one event. In such case, all date tags are recorded for later sanity checks, but the date tag appearing 昀椀rst is chosen as the dating of such event. Allowing an event to have multiple date-tags is useful when parsing chronicle entries dated with a range of dates. For example, an entry such as `between <date A> and <date B>, there was a heavy rainfall which influenced the harvest in a bad way' is not split into two, as long as both date tags appear on the same line.
The chronicles that were digitised by the DBNL pose a challenge here, since their lines do not match the actual lines in the original manuscript. Instead, their lines are arti昀椀cially created when converting them to page-XML in order to upload them in Transkribus. Since chroniclers were very inconsistent in their use of punctuation, these lines sometimes have the length of one paragraph. This is not a problem as long as such a chronicle entry corresponds to a single dated event. However, in some cases one document can consist of multiple dated entries referring to di昀erent events. Luckily, the vast majority of the documents we used to 昀椀t the novelty signal have a single date-tag (20,372 out of 22,516 documents) and thus correspond to a single dated entry. Furthermore, whether a document has a single date-tag or multiple does not seem to predict novelty scores very well (seeFigure 4). In summary, the sanity checks we conducted did not uncover a consequential bias to our analysis, resulting from our document segmentation strategy.
Annotators have marked words that span multiple lines with a special character. First, we concatenate these into a single word and remove all the other special characters used by annotators. Next, unique IDs are assigned to documents, which can be used interchangeably in both the annotated and corrected corpus.
From here on, the processing steps di昀er for both corpora to re昀氀ect the task they are used for. The annotated corpus is used for training of theTop2Vec model and should as such o昀er both su昀케cient diversity and number of texts. For this reason, we only exclude documents shorter than 50 characters (mostly OCR artifacts, or non-documents, such as page numbers). In total, 116,023 documents are passed on toTop2Vec training with the average length of 577 characters (SD=1260).
On the other hand, the corrected corpus is used to 昀椀t the novelty signal and should in turn be reliably transcribed and dated. First we attempt to capture events that can be dated up to a day (i.e. having a fully speci昀椀ed date tag in a YYYY-MM-DD format, as opposed to e.g. YYYY-MM). This undoubtedly has an e昀ect on the results, since events that are not connected to a speci昀椀c date by the chronicler, but instead to e.g. a month, are excluded. This is further discussed in the 昀椀nal section of the paper. Second, we exclude documents shorter than 50 and longer than 5000 characters. With this additional upper limit on document length we attempt to exclude events, which contain verbatim copies of o昀케cial documents and non-events, such as chronicle appendices. These very long documents are outliers in both corpora (less than 1% of the documents in thecorrected corpus are longer than 5000 characters) and a majority of them contain a single date-tag, meaning they are not a concatenation of multiple short events. Finally, only documents dated in the period of interest (years 1500 through 1820, inclusive) are kept. In total, 36,147 documents are passed on to the next step of novelty detection with the average length of 525 characters (SD=635).
We use a Top2Vec model to create both document representations and topics1[]. This model is based on the assumption that many semantically similar documents are indicative of an underlying topic. Consecutively,Top2Vec creates jointly embedded document vectors and word vectors using doc2vec, it creates lower dimensional embeddings of document vectors using UMAP, and it 昀椀nds dense areas of documents using HDBSCAN. For each dense area, the centroid of the document vectors is then calculated, which is assumed to be the topic vector. Finally, it searches for the n-closest word vectors to the resulting topic vector, in order to create a topic.
An important di昀erence between Top2Vec and traditional bag-of-word topic modeling methods such as LDA, is that the semantic embedding used in the 昀椀rst method has the advantage of learning the semantic association between words and documents. We consideTrop2Vec therefore a more suitable method, since the corpus at hand contains large spelling variation, for which the semantic embedding approach can serve as a solution. Furthermore, LDA models topics as distribution of words, which are then used to recreate the original document word distributions with minimal error. This o昀琀en necessitates uninformative words which are not topical to have high probabilities in the topics since they make up a large proportion of all text. A stopword list can be used to solve this problem, but expanding the list in order to get rid of these non-topical words can be a never ending iterating process.Top2Vec does not need a stop list, because high frequency words that occur in all documents will not be particularly close to any topic vector and thus not dominating in any topic.
We train a model on all primitives annotated that are longer than 50 characters. The trained model contains 426 topics, but we reduce it to 100 topics (a昀琀er using the elbow method to 昀椀nd the optimum number of topics, see Figure 5), using the hierarchical topic reduction function inTop2Vec, which 昀椀nds the representative topics of the corpus by iteratively merging each smallest topic to the most similar topic until the number of 100 topics is reached. The model provides two relevant outputs, which are the vector representations of the documents, and the cosine similarities of the documents towards the topic centroids. Concerning the latter, it is important to note thatTop2Vec, being a geometrical model, di昀ers here from a probabilistic model such as LDA. The ‘weight’ is therefore not a probability between 0 and 1, but the cosine similarity between a document and a topic, which is a value between−1 and +1.
As was mentioned earlier, there is a sharp di昀erence in the number of primitives across years in our corpora. In order to alleviate this problem, we pick a single ‘prototypical’ document for each day if there are multiple documents tied to that day. To acquire prototypes, we 昀椀rst group doc2vec document embeddings (acquired in the previous step) into daily subsets. For each subset, we then calculate pair-wise distances between embeddings. The embedding with the lowest average distance to all the other in the subset is then picked as the prototype. The distance metric used is cosine distance:
Hereby we aim to capture the document that is most similar to other documents in a daily subset. For example, if multiple documents refer to the same event on the same day, only one will be picked to represent it. Furthermore, this step allows us to regularize the interval of = 1 − A ⋅ B
‖A‖‖B‖ measurement (time elapsed between datapoints). Regular intervals are important for choosing the window ( ) parameter in novelty detection, as well as interpreting the resulting novelty values; A昀琀er choosing prototypical events, a primitive with high novelty can be considered novel in the context of or more days, and not just an eventful a昀琀ernoon with records (e.g. Stadtholder Willem V visiting Purmerend).
Furthermore, a sanity check in which we did not choose prototypical primitives revealed that the novelty peaks remain practically unchanged. It is therefore very unlikely that the peaks are driven by the picking of irregular documents as prototypes.
The following measures are calculated on 300-Ddoc2vec embeddings of the chosen prototypical events, ordered by date. First, embeddings are turned into a probability distribution using the so昀琀max function: We then proceed calculate novelty, transience and resonance. Withnovelty, we refer to an event (Ā)’s reliable di昀erence from past events (Ā−1), (Ā−2), …, (Ā− )in window : ( )ÿ=
ÿ ÿ ∑Ā=1 ÿ Ā ℕ (Ā) =1 ∑ þ=1
( (Ā)∣ (Ā−þ)) ℝ (Ā) = ℕ (Ā) − (Ā) and resonance as the degree to which future events (Ā+1), (Ā+2), …, (Ā+ )conform to an event (Ā)’s novelty: where is the transience of (Ā): This model for novelty and resonance was originally proposed in3[], but here we use the symmetrized and smooth version with the Jensen-Shannon divergence ( ): where = 1 ( (Ā)+ (ā))and 2 ( (Ā)∣ (ā)) = 1 ( (Ā)∣ ) + 1 ( (ā)∣ )
2 2
signi昀椀es Kullback–Leibler divergence [
(Ā)
(Ā)
( (Ā)∣ (ā)) =∑ ÿ log2 ( ÿ )
ÿ=1 ÿ(ā)
(6)
(7)
In this case, Jensen-Shannon divergence is preferred over Kullback–Leibler divergence for a
number of reasons. First, we maintain that allows us to relax assumptions about the
temporal order of observations, as it is a symmetric metric (meaning ( | ) = ( | )for
probability distributions and ) [
Nonlinear adaptive 昀椀ltering is applied to the information signals because of the their inherent
noisiness [