In this paper, we study how data from the Semantic Web can be used for case studies in the Humanities. We conduct a proof of concept, using the YAGO knowledge base to study life expectancy, birth rates, and the age at childbirth over time. We also discuss the information extraction methods that we used to make YAGO su ciently complete for these analyses to work.
3 Incompleteness is a general problem on the Semantic Web: In DBpedia [
This paper is structured as follows: Section 2 discussed related work. Section 3 presents our improved methods for information extraction, which we evaluate in Section 4. Section 5 presents our case studies, before Section 6 concludes. 2
Several projects have recently started to create large knowledge bases. Among
the most visible ones are YAGO [
The idea of using data from the Semantic Web to support the Digital
Humanities is in its infancy [
In accordance with the RDF standard, YAGO stores information in the form of triples. Each triple consists of a subject, a relation name, and an object { as, e.g., in hCleopatra, wasBornin, Egypti. In total, YAGO contains 10 million entities (like persons, organizations, cities, etc.), and more than 120 million triples about these entities. The KB knows 100 relation names, which have been de ned manually. YAGO assigns each entity to one or several classes. Cleopatra, e.g., is in the class People from Alexandria. This class is a subclass of the class Person, which is in turn a subclass of Organism, and so on.
The main part of the information in YAGO stems from Wikipedia. Every
article in Wikipedia becomes an entity in YAGO. The triples are extracted from
the infoboxes in Wikipedia and the category names. This works with a modular
architecture [
One particularity of YAGO is that it has a manually evaluated precision of 95% with respect to Wikipedia. This means that, statistically, only 1 triple out of 20 does not correspond to the facts in Wikipedia. All of our improvements have to respect this quality constraint. 3.2
The infoboxes of Wikipedia do not mention the gender of a person. Therefore, earlier versions of YAGO did not have gender information. Gender was added only in YAGO3, and it was extracted from the occurrence of pronouns on the page. The GenderPronounExtractor counts the number of occurrences of \he" and \she" in the articles. If the number of \he" is at least twice the number of occurrences of the word \she", and if the number of occurrences is at least 10, the gender is assumed to be male (and vice versa). This worked well, but it had a rather low coverage. Only 61% of people had a gender.
We improved this method by making use of the gender-speci c categories in Wikipedia. For example, Cleopatra is in the category Female people from Alexandria. We wrote the GenderCategoryExtractor, which considers every article x, and produces the fact hx, hasGender, femalei if the article is in a category that contains the substring female (analogously for the male categories). This works well, but it still has a low coverage.
We improve upon this as follows: We wrote the GenderNameExtractor, which
collects the rst names of all people with a known gender. The same rst name
may be associated to people of di erent gender { either because the name is used
for both males and females, or because of errors in Wikipedia or the extraction
process. Our goal is to determine whether is associated to one gender in the
majority of cases. Let us say that our sample for name has a proportion of p%
males. We want to know what is the proportion of male people with the name
in the real world. For this purpose, we run a statistical test: We assume that
our set of people with rst name is a sample from the real world. Then we
generalize the proportion of males in our sample to the proportion of males in
the real world. We use the Wilson estimator for this purpose, with = 5%. Only
if the lower bound of the Wilson interval is at least 95%, we assign the name
to the male gender (analogously for female). The values of and 95% are those
that the YAGO evaluation [
All of these three extractors produce sets of facts in the YAGO framework. These are then collected by a fourth extractor, which assigns at most one gender to each person, giving priority to the GenderCategoryExtractor, followed by the GenderNameExtractor and the GenderPronounExtractor. 3.3
YAGO harvests the birth dates and death dates of people from two sources: From the infoboxes of Wikipedia and from the categories of Wikipedia. The infoboxes contain a list of attribute-value-pairs, of the form date-of-birth = Jan 8, 1935. As for the categories, there exists one category per birth year (e.g., 1935 births ). YAGO uses the date from the infobox where available, and defaults to the year from the category otherwise.
The problem with this approach is that the extraction quality is higher for the categories than for the infoboxes. This is due to more varied date formats in the infoboxes (01/08/1935, 08/01/1935, 8 Jan 1935, etc.). Therefore, we modi ed the respective extractor as follows: If only one birth date is available, that date is used. If the year from the infobox and the category coincide, the date from the infobox is used. Otherwise, we use the year from the category (and abandon the date from the infobox). We proceed analogously for death dates. 3.4
YAGO extracts the birth place, death place, and place of residence from the infoboxes of Wikipedia. The problem with this approach is that the data in the infoboxes is very sparse. However, some categories give away the nationality or the place of residence of a person, as in Egyptian queens regnant.
Therefore, we have written a new extractor (CategoryLocationExtractor ) that harvests also the categories of Wikipedia. We rst compiled a list of demonyms { in part from the Wikipedia list of demonyms, and in part from the lists of empires on Wikipedia. Our extractor then scans the categories of a person x, and counts the number of times each country's demonym appears. For every country y that appears most often, we create a fact hx, livedIn, yi. Our rationale is that the categories can neither determine the birth place nor the nationality reliably, but that they can at least indicate a place of residence. 4
In this section, we study the e ect that our new extraction methods have. We
consider two axes: Precision and coverage. For precision, we took a sample of
100 facts per method, and checked them manually against Wikipedia, as we
usually do for YAGO [
We also report the coverage in DBpedia, the other big knowledge base extracted from Wikipedia. These numbers have to be taken with a grain of salt: First, there exist di erent versions of DBpedia, one for each Wikipedia language. We take here the English one (which is the largest). YAGO, in contrast, extracts from 10 Wikipedias. Second, the precision of DBpedia is not known. We did not want to make statements about the quality of DBpedia in place of its authors. Thus, our results should not be understood as a direct comparison.
Table 1 shows our results. YAGO contains 2,284,927 people. Our precision values are very good. The 2% wrong results for genders were exclusively due to type errors, not to the extraction itself. The extraction of residences also su ered from that problem. In addition, it produced 6% of anachronistic residencies (such as German Empire instead of Germany ). We counted them as correct for the purposes of our study. The precision of birth and death dates is extraordinary.
In addition, our methods have signi cantly increased the number of data points for all attributes that we consider. Genders, e.g., increased by 35%. Almost three times as many people as before have a place of residence.
Extraction Birth Dates Death Dates Place of Residence Gender 4 from the DBpedia statistics Web page 5 union of birth-place, death-place, and residence from \Mapping based objects" 6 \Gender" dataset
Our rst study (Figure 1) investigates the pattern of life spans across history. It does so by plotting the average life span of males and females against time. We restricted the study to centuries where we had more than 100 men and women, respectively. Interestingly, there is no trend in the data until the 15th century, with the average age uctuating around 53 and 60 years for females and males, respectively. The e ects of the Black Death are clearly mirrored in our data: life expectancy decreases in the 13th century. Beyond that, there is an steady increase in the life span across genders during the last 500 years. We also see that women generally had a shorter life time in our data. This changes, however, in the 19th century: Women live longer than men. As our data is quite dense from the 19th century onwards, this fact is statistically signi cant in our data. 80 75 70 e ag65 n a i d e M60 55 50
1100 1200 1300 1400 1500 1600 1700 1800 1900
Year
We can also drill down into the historical life expectancy per country. Figure 2 shows 4 countries that existed continuously over the past 1000 years, together with the median age of their population (only for centuries with more than 100 data points). The gure shows that the life span has been increasing since the Black Death in the two developed countries in the sample, Italy and Great Britain. We also see a catch-up e ect: India and China have been experiencing a much larger increase in life span during the last 200 years. Another interesting observation is the take-over of Italy by Great Britain. This is mainly due to a deceleration of Italy rather that an acceleration of Great Britain, and it is thus di cult to argue that is related to industrial revolution. Today, all 4 countries have a comparable life expectancy in our data. This result has to be taken with a grain of salt: Wikipedia (and hence YAGO) contains mainly the elite population. Results may thus not generalize to the full population of a country. Our next study will shed light on that divergence. 5.2
Births per month
9 % sh8.5 % t r i b e v i ltea 8 % R 7.5 % 0 2 4
6 Month 8 10 12
Our next study tests whether the likelihood of making it to Wikipedia/YAGO is associated with the month of birth (Figure 3). For this purpose, we plotted the number of births in the US per month in Figure 3, and compared it to the number of births per month according to the U.S. birth registry7. We nd that generally, both graphs show the same peaks. However, we nd that people born in January are slightly more likely to be in Wikipedia than expected.
One potential explanation for this pattern is that our data comprises many sportsmen, who bene t from the relative age e ect8: Sportsmen born early in the year are slightly older and thus slightly more mature than sportsmen born later in the year with whom they usually compete. This makes them more successful, and hence slightly more likely to appear in Wikipedia. Indeed, if sportsmen are removed from the graph (\YAGO other"), our curve becomes more similar to the US census data. 7 From the National Center for Health Statistics: http://abcnews.go.com/Health/
Science/story?id=990641, [
A popular myth has it that there is an increase in births on full moon days9. We wanted to analyze this conjecture with our data. For this purpose, we computed the time points of the full moon over the past centuries. We restricted our analysis to 1600 to 2000, as our full moon calculation bases on the Gregorian calendar introduced in 1582. We also did point checks with known historical full moon dates to make sure our computation is correct.
We consider the two days that precede, and the two days that follow a full moon day as \full moon days". Among the 146,463 days between 1600 and 2000, we classi ed 24,793 days as full moon days. This gives a total ratio of full moon days of 16.9%. During that time period, 691,616 people were born, of which 117,081 fall on a full moon day. This gives a total ratio of full moon births of 16.9% { exactly as expected. The Wilson score interval at = 95% is 16.9% 0:1%. Thus, the full moon has no in uence on the birth rates in our data.
Additionally, we plotted the proportion of people born on or around a full moon day between 1600 and 2000 per century. Figure 4 shows our results. The great uctuation seems to indicate a relationship between the period of the moon and the frequency of births. However, this uctuation is mainly due to sparse data. As the con dence interval shrinks, the proportion of full moon births converges to the expected value of 16:9%.
0.28 0.26 0.24 9 https://www.google.com/#q=full+moon+birth
We now turn towards the age at which people become parents. Figure 5 shows the age at which people have their rst and last child during the last millennium. Again, we restricted our analysis to centuries where we had at least 100 parentchild pairs. Males show an increase in the age at which they have their rst child, moving from around 28 to around 32. For females, we see similar story, with one important di erence: for females, the increase is concentrated during the last 2 centuries. However, the age at which males have their last child is almost unchanged until 1700s, and declining since then. For females, the age at the birth of last child is unchanged. We speculate that two demographic phenomena are creating these patterns. First, women having less children by mainly postponing the age at which they have their rst child. Second, the age di erence between fathers and mothers is declining. These two forces together can create an increasing rst-child age of mothers and a declining last-child of fathers with constant last-child age of mothers and rst-child of fathers. 40 35 d o o thn30 e r a p f eo 25 g a n a i ed20 M 15 900 1100 1300 In this paper, we have investigated how data from the Semantic Web can help research in the Digital Humanities. As a proof of concept, we have used the YAGO knowledge base, one of the largest general-purpose knowledge bases on the Semantic Web, to study the life expectancy of people across di erent times; the age at rst childbirth; and the myth that full moon days see more births. We have also presented methods to improve the coverage of YAGO.
Our methods are integrated in the YAGO infrastructure, and the data we generate will be included in the next release of YAGO. The data of YAGO can be downloaded freely at http://yago-knowledge.org. A SPARQL endpoint is available at https://w3id.org/yago/sparql. We also provide the source code used for this study. The code for generating the previous YAGO is available at https://github.com/yago-naga/yago3/tree/whise2017-YAGO-before. Our new algorithms are available at https://github.com/yago-naga/yago3/tree/ whise2017. The data generated for this study is available at https://www. thomasrebele.org/projects/whise2017. The entire YAGO project is available as open source at https://github.com/yago-naga/yago3. Acknowledgments This research was supported by the grants ANR-11-LABEX0045-DIGICOSME and ANR-16-CE23-0007-01 (\DICOS").