This paper proposes a methodology for studying canonisation of people in history with digital methods. These canons are for the most part culturally determined. For a select group of people, there is no doubt that they merit the necessary attention, but there is a large gray field of 'second rate' individuals who had an impact on history of which only a small group is included in more than a footnote. This makes the attention people get from historians rather arbitrary, subjective and unacademic. Digital humanities technologies can help us to work around this arbitrariness and to get insight into the canonisation processes.
This paper proposes a methodology for studying
canonisation of people in history with digital methods.1 With
‘canonisation of people in history’ we mean the repeated
mentioning of people in any history book (e.g. a study
on British Parliament), reference work (e.g. a
biographical dictionary), newspaper, website or actual canon (e.g.
the ‘Canon van Nederland’.2 Canons are for the most part
culturally determined, rather than by the actual impact
people had in history. The example of the continuous
underrepresentation of women in history works makes this only
too clear
In this paper we take canonisation of individuals in the Netherlands as our example, but the same methodology could be applied to other countries. The rest of this paper is structured as follows. In Section 2, we introduce the phenomenon of canonisation in history and the role digital humanities can play. Section 3 discusses the different sources that could provide an answer to our question. In Section 4, we provide a breakdown of the available biographical data and tools for the Netherlands, how to make good use of them and what their limitations are. We propose a methodology for making the best use of digital methods in combination with traditional methods for canon breaking research in Section 5. In Section 6 we show some preliminary results, which is followed by our conclusions.
1All URLs in this paper were latest retrieved on 31 May 2015 2http://www.entoen.nu/ 2
Canonisation of people and events in history is an
unfortunate, but natural process. Once individuals are mentioned
and remembered in various sources, they enter the
frameworks people use to maintain their memory for a longer
period of time
One of the main challenges in addressing this problem is that identifying influential people that did not make it into the history books is a process of collecting needles from a haystack. Historians need to go through vast amounts of data that contain references to influential people and find those people that are forgotten despite being equally influential as their famous or semi-famous contemporaries. Digital methods are necessary to carry out such research in an efficient way.
The advent of the digital age has in general sparked a new
interest in frequency lists, which help us in understanding
canonisation processes. Ngram viewers can tell us the
frequency of a (combination) of words within a certain corpus
of texts over time, we can count the number of words used
by members of parliament and the kind of terms they use
and we can evoke fame rankings of people who are
mentioned in Wikipedia.3 Such lists are particularly interesting
for humanities researchers, since they give them the
opportunity to approach old topics in a different way.4 Computer
software is able to analyse much more text than any human
could ever do, which allows humanities researchers to back
up interpretations based on anecdotal evidence with actual
numbers and to formulate or test hypotheses more quickly.
With the Google Ngram viewer, based on the words in
millions of books, it is for example easy to see how the
popularity of Anne Frank rises quickly after the Second World
War.5
The creators of the Google Ngram viewer have run some
interesting experiments with their corpus
Online biographical dictionaries and Ngram viewers give
ample possibilities for investigating who became famous
and why, even when taking all the source biases and
limitations of the tools into account. It is more challenging
however, to look for the people who did not become famous,
while they were prominent enough in their own time. Even
if the data in Google Books and the KB Ngram viewer are
less discriminative than the biographical dictionaries, they
do not solve this problem. When the creators of the Google
Ngram viewer did their research on the fame of people
between 1800 and 2000 they used existing lists of people from
Wikipedia and the Encyclopedia Britannica. Even if the
lists from the Encyclopedia ‘reflect a process of expert
curation that began in 1768’
These experiments are, in other words, top-down: existing lists were used to match with records of the past. The experiment can show that certain people are not mentioned as much as one would expect, but not that certain people or events were ‘hot topics’ during a certain time, but have been forgotten since. For a complete picture, records need to be 3Wikirank: http://wikirank.di.unimi.it/ index.html; Pantheon: http://pantheon.media. mit.edu/methods
4e.g. When was the word potato used for the first time? ‘De DBNL ngram-viewer van de KB’: https://www.youtube. com/watch?v=XpMqypF46RY
5https://books.google.com/ngrams/ search for ‘Anne Frank’, on 13 May 2015.
6http://www.sciencemag.org/site/feature/ misc/webfeat/gonzoscientist/episode14/ index.xhtml queried that do not have the bias of modern records. We want to scan for any names in a wide variety of not only books, but also sources like journals, newspapers, pamphlets and archive material, and see what happens to their fame in the course of centuries. In Section 3, we will say a bit more about the potentially interesting sources to use to get a grasp on these ‘missing persons.’ 3
To map canonisation in history we need to make a distinction between the different sources we can use. There are contemporary sources (e.g. a pamphlet from 1581 scolding William of Orange) on the one hand and sources written after the death of a person (e.g. a biography on William of Orange from 1978) on the other. Similarly, there are sources with a conscious selection of people (e.g. historical sources like a biographical dictionary) and sources that do not or less consciously select (like a list of land owners). Obviously we can have both contemporary and later sources with and without a conscious selection, as can be seen in Figure 1.
The contemporary sources are needed to see how famous a person was in his or her own time. We will see in Section 6, Table 2 for example, that the politician Johan Rudolph Thorbecke was extremely prominent in the newspapers of his time. It is logical to assume that a person is often most famous in his or her own time, but the examples of the painter Vincent van Gogh and Anne Frank already show that this is not always the case. The sources published after the death of an individual show how the fame of a person developed. Even though Thorbecke remained one of the canonised figures from Dutch history, his fame declined over the years, as can be seen from the sources after his death. Obviously, for historical figures before the nineteenth century this starting point will be difficult to determine, due to the lack of sources.
It is more complex to make a distinction between sources that consciously select people to write about and sources that do not. A biographical dictionary is a good example of a source that does consciously select individuals. One of the main questions of any editor of a biographical dictionary is who is noteworthy enough to get an entry and who is not. A history book on the Dutch Revolt is already a less clear example of selection. Obviously, any historian selects the people and events he or she deems important enough to describe. The mentioning of individuals might, however, have to do with the selection of an event (e.g. presence at a certain battle) rather than with any selection of persons. This is why we consider prosopographical studies, group biographies, as good examples of sources that do not consciously select individuals. Prosopographies are quantitative studies on larger groups of people. The category a person belongs to (e.g. officers at the Council of Holland) determines whether someone is selected for the study, not the person him- or herself. Newspapers also select what they deem the most important news, but that is mostly driven by popular demand and not by historical judgments on who is influential enough to include.
The differences between these sources have to be taken into account for any historical interpretation of the results. In the following section we shall see how this already is facilitated.
To investigate who became a well-known person, who did
not, and why, we need at least the following data from as
many records as possible: names, dates and places of birth
and ‘claims to fame’ (i.e. why did or could someone
become famous?). In theory any record of people or events
could be suitable for our purpose, from medieval
chronicles to early modern newspapers, to modern school books.
For a full picture of canonisation a wide variety of sources
needs to be consulted, from each category listed in Figure 1.
Lists of famous people are strongly dependent on the kind
of medium that is consulted, as will be demonstrated in
Section 6. The enormous amount of data from these sources
could never be close-read by one person. We therefore need
digital methods to speed up our research
Resources such as DBpedia,9 a structured dataset in RDF
based on the data in Wikipedia, offer similar possibilities
for group analyses of ‘famous’ people. The advantage of
these datasets over biographical dictionaries is that they are
bigger, dynamic, more inclusive and edited by ‘the crowd’
rather than by a selected group of editors. Wikipedia does
particularly well in providing reliable basic data on
individuals. One of the disadvantages is that DBpedia and
Wikipedia have clear biases as well, which are more
often grounded in ‘Geek hobbies’ than in academic tradition
In this section, we propose a method for fruitful
computational analysis of canon formation with digital historical
data. As mentioned above, Ngram viewers are suitable for
‘top-down’ research on canons, when you know which
people you are looking for. We want to combine this approach
with a bottom-up approach, where the starting point is not
an existing list of names, but all the names from as many
resources as possible from all categories as described in
Section 3. This way we can also find whose fame did not last
for centuries and formulate ideas on why this is the case.
Another identified problem with Ngram viewers is that they
provide little context and provenance information.
Especially for a historian, it is important to know where
information came from, to check the reliability and to see the
context
For this paper we performed step 1 and applied a basic approach to address step 3.
6
In this section we will discuss the results from a top down
approach for investigating who is most famous in Dutch
history. Since any existing fame list would do as a
starting point, we took the top 25 of the Dutch TV elections of
the ’Grootste Nederlander’ (Grandest Dutch person).18 We
then ranked them basing ourselves on the Google (books)
Ngram viewer (for English), the KB Ngram viewer,
taking the words from Dutch newspapers, the DBNL Ngram
16IRIs are generalizations of URIs that support Unicode.
17Note that there is no predefined ontology, which makes this
a different task from standard named entity disambiguation as in
18http://nl.wikipedia.org/wiki/De_ grootste_Nederlander viewer, containing words from mostly literary texts and Google Ngrams for Dutch, which contains all words used on the Internet at the end of 2008. With these sets we have sources from historiography, the news, cultural texts and the Internet, which together should provide a rather balanced set of sources with much and less selection, both from the period during and after individuals’ lives. We ranked the individuals by their highest score in one year, since for the limited scope of this paper it would go too far to calculate a balanced average for each individual. We faced several challenges in identifying the right people. The spelling of names is possibly the biggest issue here. Before the nineteenth century there was no standardised spelling of names, which results in many varieties in not only contemporary sources, but also in modern works. Even if a particular name is usually spelled the same way, a bad OCR quality could still give a bias in the results. The options to use wildcards in the viewers to catch all variations often are very limited.
Another problem is caused by people with the same name. William the Silent, number two in the elections (see Table 1), is most commonly known as William of Orange. The hits we receive for ‘William of Orange’ in the Google Ngram viewer however, may refer to the leader of the Dutch revolt (y1584) we are looking for, but also to his greatgrandson, the later King of England (y1702), number 72 in the TV elections. Pollution with instances of the king of England could be especially significant in the English corpus of Google books. We therefore only used his nickname ‘William the Silent’ in this corpus. Despite the significant reduction in hits, he still ranks number 1 in Google books, which further justified our decision.
Identifying the humanist scholar Desiderius Erasmus poses a problem because he is known as ‘Erasmus’. Dropping his first name would lead to many additional hits from other people and Google NGrams for Dutch does not even facilitate searching for unigrams. The same applies to the philosopher Baruch de Spinoza. A quick search in the World Biographical Information System19 shows us that while there are 789 hits for Erasmus, there are ‘only’ 8 hits for Spinoza (and most of them refer to the correct and the same person) indicating that the risk of pollution is lower. Still, results in Google Ngrams seem significantly inflated for the unigram Spinoza, giving him an extremely high score in 1883. The year 1883 does not have a high score when searching for bigrams of ‘Baruch Spinoza’, or trigrams of ‘Baruch de Spinoza’, which strongly suggests that too much pollution occurs when the first name is dropped. We therefore added the results for ‘Baruch Spinoza’ and ‘Baruch de Spinoza’, whilst knowing the score does not reflect all references to him.
There also are people who are known differently during their lives, such as members of the royalty. We had to search for both princess and queen Juliana and princess and queen Wilhelmina to obtain the best result. For widely known people like them this problem can be circumvented quite easily, but in other cases specific domain knowledge 19http://db.saur.de/WBIS/basicSearch.jsf The system hosts biographies on 6 million people from 58 biographical archives all over the world.
Elections 2004
Pim Fortuijn Willem van Oranje
Willem Drees Antoni van Leeuwenhoek
Desiderius Erasmus
Johan Cruijff Michiel de Ruyter
Anne Frank Rembrandt van Rijn Vincent van Gogh
Total NGram viewers Koningin Wilhelmina Willem van Oranje
Koningin Juliana
Vincent van Gogh Rembrandt van Rijn
Anne Frank Johan Thorbecke Christiaan Huygens Desiderius Erasmus
Prins Claus
is needed to find all instances. To give just one example:
Dutch treasurer Vincent Cornelisz from the first half of the
sixteenth century was very famous in his time, but is
currently unknown to a wide audience. In history books he is
not only referred to as Vincent Cornelisz, but also as
Vincent van Mierop (a name which was used for the first time
by his son, not by him), or as Vincent Cornelisz van Mierop.
In records of his own time, he was so well known that often
he was simply referred to as master Vincent, which
ironically means that the fame in his own time causes a problem
in tracing his fame in our time
As mentioned in Section 5, it is relatively easy to identify names with tools for Named Entity Recognition. For this particular study, we use a highly simplistic but effective pattern-matching approach. We select combinations of words that start with a word that starts with a capital (e.g.
Willem) and end with a word that starts with a capital (e.g.
Oranje), which works fine for Dutch (but would be quite useless for German that capitalises all nouns). Because both the first and last word must start with a capital letter, we avoid the inclusion of words that start the sentence.20 The algorithm allows for two sequential lower case words within the name, since it is customary to write prepositions and determiners in Dutch names in lower case when they are preceded by a first name or initials. The algorithm can thus capture names such as Johan Derk van der Capellen tot den Pol, but no names where three lowercase words follow each other which are extremely rare in Dutch.
For our particular use case, we primarily aim for recall, because (1) historians can immediately filter out the invalid patterns found by our approach and (2) bad patterns are often singletons in the corpus having no or little influence on the top and middle of our frequency based lists.
For these reasons, precision can be as low as 5% or 10%
20Names such as Willem II are identified, because the sources use Roman capital letters to add numbers to nobility with the same first name. and the approach will still serve its purpose (though higher precision does make the historian’s job easier). Our basic pattern-matching approach is thus preferable for this particular research over the more sophisticated machine-learning approaches that have higher precision, but lower recall.
We tested our method on the data of the Biography Portal of the Netherlands, an aggregated dataset of 23 different sources, all with their own limitations and biases.21 A biographical dictionary in itself is a ‘canon’ of noteworthy people and will therefore not reveal many ‘forgotten’ people. The Portal nevertheless provides a suitable dataset to try out our methodology. It provides a large volume of descriptive texts in Dutch and the output of our algorithm will reveal what person names occur most in these texts (in their own and in other people’s biographies) and thereby showing us a measure of fame after all. The principle of applying Google Ngrams for Dutch our method does not differ from applying it to a set that did Marco van Basten not apply any form of selection.
Anne Frank With our approach we could easily get the number of ocKoniPnigminFWoritluhiejnlmina currences of entries such as Willem van Oranje (William of Johan Cruijff Orange). The fact that we also got results from ‘Tweede Toon Hermans Kamer’ (Dutch parliament, 882 hits) ‘Den Haag’ (The Koningin Juliana Hague, 830 hits) and ‘Staten van Holland’ (States of HolWillem van Oranje land, 420 hits) shows an interesting overall bias towards poVincent van Gogh litical history, but can be easily discarded for our purpose Prins Claus here. You do not need to be a domain expert to easily see that these expressions do not refer to people.
Named entity disambiguation is more problematic. After discarding the false hits we have Willem I, Willem II, Willem III, Willem IV and Willem V ranking in the top 10 of our list, but unfortunately there have been many counts, dukes and stadtholders over the centuries who go by that name and title. A problem of a different nature is that we have Willem I, Willem van Oranje and the prins van Oranje ranking high, which could all refer to the same person: William the Silent (of Orange), the number 2 from the TV elections and the overall ranking in Table 1. Hits such as ‘Van den Bergh’ also causes identity problems, since without the context we cannot see which Van den Bergh this is, or even if he or she is an actual historical person or just a historian who is cited often. Some of the results are quite telling, however. We are quite sure that ‘Karel V’ will almost always refer to emperor Charles V (and perhaps a few times to the fourteenth century French King) and that Frederik Hendrik and prins Maurits refer to the famous sons of William the Silent. Domela Nieuwenhuis must refer to the social anarchist Ferdinand Domela Nieuwenhuis, since he has quite a unique name.
In a first attempt of named entity disambiguation we investigated the possibilities of applying time constraints based on metadata and temporal expressions in the text. This way count Willem II (thirteenth century), stadtholder Willem II (seventeenth century) and king Willem II (nineteenth century) would be easily separated.
We implemented a basic approach that tackles the time constraint of identity, which is based on the idea that people can only personally interact with someone who was alive at the same time as they were. Because this is the case, we as
BP first results
Willem I
Willem III Prins van Oranje
Karel V Willem II
Willem V
Frederik Hendrik Domela Nieuwenhuis
Willem IV Prins Maurits
BP second results Willem I (1772) Karel V (1500) Willem II (1792) Willem III (1650)
Willem V (1748) Domela Nieuwenhuis (1846)
Frederik Hendrik (1584)
Willem III (1817) Lodewijk Napoleon (1778)
Willem IV (1711) sume that in the typical case, people who are mentioned in someone’s biography will be a contemporary of the biography’s subject. In order to establish which mentions refer to the same person, we extracted the date of birth and date of death from the metadata of the biographies in our corpus.
While going through the corpus to identify names, we only merged names when the lifespan of the subjects either overlapped or were maximum 50 years apart from each other.
This baseline assures that, if the reference in the text itself is not about the far past or future, it is at least possible that the texts refer to the same person.
Because there may be people alive at the same time who have the same name and 50 years offers quite a range, the approach does not offer any guarantees that references to different people are not combined, but it helps to solve some of the clearer cases where sources do not talk about the same person. It solves, for instance, the issue of high nobility with the exact same name. They are either from a different era altogether, or they have a different number behind their name.
The results of this approach are quite promising. Table 4 shows that while we previously were not able to distinguish between Willem III, the nineteenth century king and Willem III the seventeenth century stadtholder, we now have them listed as two different individuals. It also shows that Willem I does not refer to William the Silent at all, as one may expect from the lists from our top down approach, but to nineteenth century king Willem I. Looking at the tables it seems that the Biography Portal of the Netherlands, and then most likely especially the two biggest dictionaries included in there from the nineteenth and early twentieth century, are strongly biased towards the House of Orange. Further research might show that many people were included in the dictionaries because of their link to king Willem I.
By refining this method, for example by automatically merging similar instances like ‘Willem I’ and ‘Koning Willem I’, and by applying it to a larger and a wider variety of datasets we would become closer to seeing canonisation patterns than traditional research could have ever brought us.
In an attempt to trace the ‘forgotten’ individuals we made a list of the people who do get mentioned frequently in the texts from the BP, but who do not have a biographical entry of their own. The results of this exercise were interesting enough, but do still involve quite a lot of handwork from the historian. Many people in the list we generated did have their own entry after all, but are mentioned in a slightly different way. Politician P.W.A. Cort van der Linden, for example, is often mentioned as Cort van der Linden (16 times) and similar issues occur with many other politicians from the nineteenth and twentieth century. Moreover, some individuals are known under various alternative names. For instance, sixteenth century duke Karel van Gelre is listed as Karel van Egmond.
The people who are mentioned most frequently in the texts and who really do not have their own biographical entry are listed in Table 5. We find an important religious figure, a communist philosopher (probably mainly thanks to the biographical dictionary on socialists included in the Portal), no less than eight French rulers, an English king and a German emperor in the top 12. It does not bring us closer to the forgotten people in Dutch history, but does show a clear connection of Dutch elites with French royalty (or a bias in the dictionaries towards France or people involved with France). We also encounter the previously mentioned problem of how to identify people who are mostly known with one name. To detect Erasmus in the Ngram viewers we had to search for Desiderius Erasmus. In Table 5 we see 15 mentions of Napoleon Bonaparte, while there will be many more for just Napoleon. To find them, however, we would have to expand our algorithm to include one word instances as well, which would result in too much noise for our analyses for this basic version of our algorithm.
To trace the individuals who were noteworthy in their own time, but are forgotten in history, we are more likely to be successful when analysing sources with a semi-conscious selection mentioned in Figure 1. We applied our method to a sample of 99 historic Dutch newspaper texts provided by the Koninklijke Biobliotheek.22 The sample is too small to provide indications of ‘forgotten people’, but the outcome of this test shows that our method can be applied successfully to these articles. The outcome furthermore confirmed our observation based on data from the BP that phrases that do not correspond to a name generally occur only once and therefore do not form a hindrance for the historian, given that a single mention does not point to (contemporary) fame.
7
In this paper we addressed the importance of research on canon formation in historical research. Before the advent of digital technologies and the availability of digitised data, this could only be done tentatively. We have shown that despite many methodological and technical problems, there is a decent amount of data available and there are tools that facilitate group analyses of famous people.
In section 5, we proposed a method to complement a topdown approach of analysing people still famous now with a bottom-up approach, which gives more room for unbiased selection, context and provenance of the data. The basic 22http://lab.kbresearch.nl/get/Downloads means to carry out such research are available. Even though methodologies for task 3) are still in a preliminary stage and the work in 4) and 5) still is labor intensive, the possibilities provided by digital humanities make this research feasible. In section 6, we discussed some difficulties in applying a top-down approach and have also discussed the first results of a bottom-up approach. A close collaboration between historians and computer scientists is a requirement to make such research successful, especially in the named entity disambiguation. Expert domain knowledge combined with complex algorithms are needed to match as many individuals correctly as possible and to signal false positives. Eventually such exercises can help us to explain why some people only get 15 minutes of fame and others live on in memory over centuries.
The approaches we presented in this paper are relatively basic. We explained that this is not an issue for named entity recognition, because precision is of minor importance for the historian investigating canonisation. We plan to experiment with alternative versions of the algorithm including a version that can handle single names such as Erasmus and Napoleon. However, given that our basic algorithm already provides results that yield interesting results, future work will mainly focus on better disambiguation. We expect that standard methods for named entity disambiguation are not the most suitable for this task and data, because they tend to make use of the content words used in the text and address a wider range of named entities than just people. We therefore expect most from a domain and target entity specific approach that combines frequency of the first and last name, information about time and place, as well as social networks.
The most important next step, however, will be to apply the methods outlined in this paper to new datasets that also provide a contemporary perspective and/or use semi-conscious selection. Contemporary sources play a vital role in identifying people who were famous and fell in oblivion thus providing the necessary means to compare and identify what aspects contribute to canonisation once initial fame is achieved.
This work was supported by the BiographyNet project http://www.biographynet.nl (Nr. 660.011.308), funded by the Netherlands eScience Center (http://esciencecenter.nl/). Partners in this project are the Netherlands eScience Center, the Huygens/ING Institute of the Royal Dutch Academy of Sciences and VU University Amsterdam. We would like to thank Dr. Ronald Sluijter for his insightful comments on an earlier version of this paper. All remaining errors are our own.
9