=Paper=
{{Paper
|id=Vol-1273/paper1
|storemode=property
|title=Spatializing a Digital Text Archive about History
|pdfUrl=https://ceur-ws.org/Vol-1273/paper1.pdf
|volume=Vol-1273
|dblpUrl=https://dblp.org/rec/conf/giscience/BruggmannF14
}}
==Spatializing a Digital Text Archive about History==
https://ceur-ws.org/Vol-1273/paper1.pdf
Spatializing a Digital Text Archive about History
A. Bruggmann, S.I. Fabrikant
University of Zurich, Department of Geography, Zurich, Switzerland
{andre.bruggmann,sara.fabrikant}@geo.uzh.ch
1 Introduction
The amount of digital text data available in online libraries has risen dramatically in
recent years. GoogleBooks or the Universal Digital Library (UDL) initiatives illus-
trate this impressively. The rapid evolution of vast digital text data archives has
spurred the growth of an interdisciplinary Digital Humanities (DH) community, as [1]
puts it, the once inaccessible has suddenly become accessible. Researchers in the
humanities and social sciences have recognized the big potential digital text archives
might offer to gain new insights on long-standing research questions. Especially in-
teresting are unstructured or semi-structured digital libraries in this context, as text
documents have been central to the humanities and social sciences long before digiti-
zation. Along these developments, the need for automatically extracting new
knowledge from text corpora using advanced data mining methods and information
reduction techniques has risen as well. Text corpora also bear exciting research ave-
nues for spatially aware disciplines and research fields, including geography,
GIScience, and the interdisciplinary geographic information retrieval (GIR) commu-
nity. This is because text documents often contain explicit and implicit spatio-
temporal and thematic information, which can be automatically extracted, reor-
ganized, visualized and analyzed for knowledge generation.
We would like to introduce our interdisciplinary research project for discussion at
the workshop which offers one avenue for bridging the gap between the digital hu-
manities and GIScience. We combine methods from geographic information retrieval
(GIR) and geovisual analytics (geoVA) in order to gain new insights from a digital
dictionary about the history of Switzerland [2]. In own prior work, we illustrated how
spatio-temporal analysis methods can be applied to a history text archive and how
long-standing geographic principles (e.g., Tobler’s law) might be verified with data
not typically employed in GIScience [3,4]. For the workshop, we present preliminary
findings of this spatio-temporal approach, and raise future work ideas for discussion.
We are particularly interested in incorporating sentiment analyses in our work, in
order to assess how (historical) places are referred to in texts over time. In our
transdisciplinary approach we aim to create a geographic information observatory.
We illustrate this by combining well established methods in GIScience to mine a typi-
cal data source for social science and humanities researchers, and by incorporating
methods typically employed in social science (e.g., sentiment analysis) to expand the
methodological toolkit in interdisciplinary GIScience research dealing with multivari-
ate (text) data. Below we sketch current ongoing work as a starting point for discus-
sion.
2 Space and Time in the HDS
2.1 Methods
We chose the Historical Dictionary of Switzerland (HDS) as prototypical text data
source for our approach, as it represents a common example of an online digital text
archive in the humanities. It contains explicit and implicit spatial, temporal, and the-
matic information in semi-structured text documents, in this case, about the history of
Switzerland. The dictionary is multi-lingual (i.e., German, French, and Italian), and
consists of 36,188 articles, categorized in thematic contributions (e.g., events), geo-
graphical entities (e.g., municipalities), biographies and articles about historically
important families. In this version, there are no possibilities to browse or query the
articles by space, time or theme.
In a first step we automatically retrieved spatial and temporal information from the
German version of the HDS. Other language versions will be considered in future
work. We employed the approach presented in [5] to automatically extract spatial
information from the articles. This resulted in 169,094 toponyms (e.g., cities, single
objects, forests, rivers and lakes, mountains) stored in a database. Next, we automati-
cally retrieved temporal information (e.g., dates such as 07/06/1856 or time periods
such as 19th century) from the HDS by employing Heideltime [6]. As a result,
510,357 temporal annotations were additionally stored in the database.
We generated graph models to conceptualize co-occurrence relationships of the 40
most often mentioned toponyms in the HDS articles over time, following the ap-
proach proposed by Hecht and Raubal [7]. We chose centuries as the temporal unit of
analysis, and thus aggregated extracted temporal annotations to centuries (e.g., the
date 07/06/1856 is a member of the category 19th century). Each article was assigned
century weights, according to the frequency of respective temporal annotations occur-
ring in the articles (e.g., article A: 19th century: 0.5, 20th century: 0.3, 21st century:
0.2). We used this temporal article weighting scheme to compute the strength of topo-
nym relationships by summing up the weighted relationships in articles where the two
toponyms co-occur. In other words, the more often two toponyms co-occur in articles
with a high percentage of temporal annotations categorized as 19th century, the higher
the weight for their relationship in the 19th century. Finally, we visualized the rela-
tionship graphs in a series of network visualizations, following the spatialization
framework by Fabrikant and Skupin [8]. A more detailed description of the approach
is presented in [4].
Fig. 1. Toponym relationship from the 19th to the 21st century [4] (map data source: swisstopo,
http://www.swisstopo.admin.ch/)
2.2 Results and Discussion
In Figure 1 three network visualizations are shown, each depicting the structural most
salient relationships (i.e., using a minimum spanning tree) between toponyms in Swit-
zerland over the last three centuries. The graphs were visualized with the Network
Workbench [9] using the GEM layout algorithm to avoid edge crossings. Toponyms
that have strong relationships in a specific century are visualized closer together on
the network, and are connected with a larger edge than those that have weaker rela-
tionships. The size of the nodes represents the importance of a toponym (i.e., its
strength) in the network, calculated by summing all weighted relationships with all
other toponyms in the network; the larger the node, the higher its importance. We also
ran the Blondel community detection algorithm on the complete network in order to
delineate toponym clusters which separate densely connected toponyms within a
community from weakly connected toponyms outside a community [10]. The same
information is depicted geographically on a map of Switzerland in Figure 1, with the
20 most frequently occurring toponyms labeled for reference.
By analyzing the network structures in Figure 1, the increasing degree (i.e., directly
connected nodes) of Zürich (the financial capital) compared to Bern (the political
capital) becomes salient over time. In the 19th century, the degree for Zürich and Bern
is almost the same (i.e., 14 and 13), but in the 21st century Zürich’s degree increases
to 15 and Bern’s degree decreases to eight. Furthermore, Tobler’s [11] first law of
geography (“Everything is related to everything else, but near things are more related
than distant things”) is evident in the community structure over time. The communi-
ties form contiguous spatial clusters in the maps in each time slice, except for the
green cluster in the 19th century which divides the blue cluster into two parts. Further,
we observe a merge of the red cluster in the Italian speaking part south of the Alps
(i.e., Lugano, Locarno, and Bellinzona) with the German speaking blue cluster in the
north of the Alps in the 21st century which could be due to the opening of the
Gotthard road tunnel in 1980 that connects Southern Switzerland with Northern Swit-
zerland. Further results and a detailed discussion is presented in [4].
3 Next Steps
At this point, we only considered the spatial and the temporal information contained
in the HDS. In ongoing and near future work we wish to focus on extracting and ana-
lyzing thematic information buried in the HDS corpus, and combine this with the re-
organized spatio-temporal data. Below we sketch ideas on how to achieve this, and
for which we would like to get feedback at the workshop.
3.1 Thematic Analysis
In order to explain found relationships between toponyms, we will study the topicality
of articles that connect toponyms in the HDS in more detail. One possible approach
for this is Latent Semantic Analysis (LSA) [12], including Topic Modeling [13]. We
will employ the Text Visualization Toolbox (TVT) in MATLAB [14] for LSA. We
aim to again automatically group articles with similar thematic content by using the
Blondel community detection algorithm, as mentioned earlier. Automatic labeling of
uncovered clusters could be achieved by the tf-idf method, for example. The joint
visualization of how thematic clusters in toponym relationships might change over
time could further help to explore reasons why toponym relationships might have
changed over time. A first conceptual approach how one might visualize this in a
(static) network is illustrated in Figure 2 below.
Fig. 2. Themes associated with the toponym relationship Zürich and Genf
A pie chart is placed on the edge linking the toponym Zürich with Genf. Distinct color
hues represent the thematic groups and the size of the pie chart sectors reflects the
importance of a specific theme for this toponym relationship.
We will have to face various challenges to implement the illustrated approach
above and we would like to raise them at the workshop for discussion. Firstly, with
Topic Modeling, for instance, one needs to determine the number of extracted topics a
priori. One possible way to do this is to match topic numbers with currently available
themes in the HDS. Another possibility might be to test the performance of different
Topic Modeling solutions (i.e., with varying number of topics) using the perplexity
measure [15]. Second, the (automatic) labeling of the extracted topics will be a major
challenge. Even though we have already worked on methods how to label and distin-
guish article groups thematically [16], most methods that we have tested so far still
require time consuming manual cross-validation and human interpretation. Third, the
more toponyms one aims to visualize in a network, as shown in Figure 2, the more
difficult the network might be to read and interpret. One solution to mitigate this, is to
adopt a cyclic user-centric design approach, that is, to create different visualization
solutions, discuss them with historians, and adapt them accordingly.
3.2 Sentiment Analysis (SA)
In a further step, we wish to employ sentiment analysis to the HDS. One the one hand,
we aim to get a better understanding about how (historical) places in Switzerland are
described by Swiss historians. While one would not expect to detect an authors’ or
editors’ positive or negative sentiments towards a specific topic in a scholarly lexicon,
one could expect that the content of an article, for example, about an infectious dis-
ease (e.g., Cholera) should be identified and classified as negative with sentiment
analysis, as it might contain many words and expressions which are commonly used
in such life-critical contexts (e.g., death). Further, we envision to study the changes of
sentiment towards a particular place over time, by incorporating the re-organized
temporal information (see Section 2).
General Inquirer (GI) is a well-known and widely used sentiment analysis tool
[17]. The GI tags and counts words in English texts by classifying them into negative,
positive, strong, active, and many other categories. Tetlock et al. [18] used GI to parse
reports of firms to quantify their financial performance. In GIScience, sentiments
expressed in Tweets has been analyzed, and visualized on maps (e.g., [19]), using
similar approaches. However, as GI only works with English texts, we have devel-
oped a prototype SA tool in Python which works similarly to the GI, but incorporates
the SentiWS dictionary [20]. The SentiWS categorizes German words into negative
(e.g., Gefahr = danger) and positive (e.g., Freude = joy) sentiments. It contains about
3,500 German sentiment words, mainly based on an automatic translation of GI’s
word categorization. Other German sentiment dictionaries such as the Berlin Affec-
tive Word List Reloaded (BAWL-R) [21] will be considered in future work.
In a first attempt, we calculated the sentiment scores for all HDS articles. Table 1,
lists HDS articles which were ranked most negatively and positively (English transla-
tions in brackets, where necessary) according to the sentiment scores (i.e., negative /
positive, etc.) which were assigned to the HDS articles by applying our own prototype
SA tool. We calculated the corresponding sentiment scores by summing up the
weights of the negative and positive words in the HDS articles, and normalized the
score by the document length. Finally, we multiplied all sentiment scores by 1000.
The category thematic contributions seems to score highest compared to the other
categories. We thus include these articles in our study for now, the remaining catego-
ries will be considered in future work.
Table 1. Top five negative and positive themes
Negative HDS articles Score Positive HDS articles Score
Articulans -19.5 Ehrschatz (legitimation fee) 5.2
Cannstatt, Gerichtstag von -17.2 Pfarrhäuser (parochial house) 4.2
Pro Libertate -15.8 Fête des Vignerons 4.0
Helvetische Annalen (Helv. Annals) -12.7 Comics 3.9
Cholera -11.5 Hülsenfrüchte (legume) 3.6
This first analysis looks promising, especially when considering the negatively scored
articles. For example, the article Articulans is mainly about banishments of people,
and about people who were sentenced to death and executed. The article Cholera
describes how Switzerland was affected by this disease. These articles contain a great
deal of negatively weighted words (e.g., Tod = death, Streit = fight, Panik = panic).
However, interpreting the positively ranked articles seems not to be as straightfor-
ward. According to the literature, reasons could be that negative information is more
salient and thus has more impact on readers, than positive information [18], and also
because of frequently appearing negations of positive words in certain contexts [22].
We will have to face various further challenges regarding sentiment analyses in
planned future work. The choice of an appropriate method in this context is of major
importance. For example, there have not been many studies about how to implement
and evaluate sentiment analyses in languages other than English (i.e., German). A
further challenge will be the effective and efficient visualization of the results. We
plan to depict historically most important places (i.e., toponyms) on a map of Switzer-
land, and visualize the potentially changing sentiments with which they are associated
in different time periods, for example, by using pie charts (i.e., different periods of
time = different sectors in the pie chart). We can then assess the valence of a place by
analyzing its association with negative or positive texts, and how this might have
changed over time. We will discuss results and visualization ideas with the historian
target group, in order to evaluate our solutions.
4 Summary
Our contribution to create a geographic information observatory incorporates a com-
bination of GIScience methods applied to data typically employed in the humanities.
We have sketched ideas and illustrated our transdisciplinary approach in combining
geographic information retrieval with geovisual analytics to retrieve, reorganize, and
visualize text information about space, time, and theme in a semi-structured digital
humanities text corpus. We also suggest how sentiment analysis applied to articles in
a history dictionary could help to reveal geographically relevant context of historical
places, and how these places are described by historians.
We further illustrate how long-standing geographic principles (i.e., Tobler’s law)
can be verified in non-typical geographic data contexts (i.e., text corpora for the hu-
manities), and how the methodological toolkit in GIScience can be expanded by in-
corporating methods, borrowed from social science and the humanities (e.g., senti-
ment analysis). We further contribute to the digital humanities by applying sound
cartographic techniques to make uncovered thematically relevant information visually
salient to a user. In doing so we hope to help historians to generate new research hy-
potheses, and to get a better understanding of complex spatial, temporal, and thematic
relationships buried in text archives.
References
1. Jockers, M.L. (2013): Macroanalysis – Digital Methods & Literary History. University of
Illinois Press.
2. Historical Dictionary of Switzerland (HDS), http://www.hls-dhs-dss.ch/
3. Bruggmann, A., Fabrikant, S.I. (2014): How to visualize the geography of Swiss history.
In: Huerta, Schade, Granell (eds.) Connecting a Digital Europe through Location and
Place. International Conference on Geographic Information Science, AGILE 2014, Jun. 3-
6, 2014, Castellón, Spain, ISBN: 978-90-816960-4-3.
4. Bruggmann, A., Fabrikant, S.I. (in press): Spatializing time in a history text corpus. In:
Proceedings of the 8th International Conference on Geographic Information Science,
GIScience (extended abstracts), Sept. 23-26, 2014, Vienna, Austria.
5. Derungs, C., Purves, R.S. (2014): From text to landscape: locating, identifying and map-
ping the use of landscape features in a Swiss Alpine corpus. International Journal of Geo-
graphical Information Science 28(6), 1272-1293, DOI: 10.1080/13658816.2013.772184.
6. Strötgen, J., Gertz, M. (2013): Multilingual and Cross-domain Temporal Tagging. Lan-
guage Resources and Evaluation 47(2), 269-298.
7. Hecht, B., Raubal, M. (2008): GeoSR: Geographically Explore Semantic Relations in
World Knowledge. In: Bernard, L., Friis-Christensen, A., and Pundt, H. (eds.) 11th AGILE
International Conference on Geographic Information Science.
8. Fabrikant, S.I., Skupin, A. (2005): Cognitively Plausible Information Visualization. In:
Dykes, J., MacEachren, A.M., Kraak M-J. (eds.) Exploring Geovisualization, 667-690.
9. NWB Team (2006): Network Workbench Tool 1.0.0.,
http://nwb.slis.indiana.edu
10. Blondel, V.D., Guillaume, J-L., Lambiotte, R., Lefebvre, E. (2008): Fast unfolding of
communities in large networks. Journal of Statistical Mechanics: Theory and Experiment,
DOI: 10.1088/1742-5468/2008/10/P10008.
11. Tobler, W. (1970): A Computer movie simulating urban growth in the Detroit region.
Economic Geography 46(2), 234-240.
12. Landauer, T.K., Foltz, P.W., Laham, D. (1998): An Introduction to Latent Semantic Anal-
ysis. Discourse Processes 25(2&3), 259-284.
13. Steyvers, M., Griffiths, T. (2007): Probabilistic Topic Models. In: Landauer, T.K., McNa-
mara, D.S., Dennis, S., Kintsch, W. (eds.) Handbook of Latent Semantic Analysis.
14. Hespanha, S.R., Hespanha, J.P. (2011): Text Visualization Toolbox – a MATLAB toolbox
to visualize large corpus of documents, http://www.ece.ucsb.edu/~hespanha
15. Blei, D.M., Ng, A.Y., Jordan, M.I. (2003): Latent dirichlet allocation. Journal of Machine
Learning Research 3, 993-1022.
16. Bruggmann, A. (2012): Netzwerkvisualisierung der Ostschweiz. Master Thesis, University
of Zurich, Zurich.
17. Stone, P.J., Dunphy, D.C., Smith, M.S., Ogilvie, D.M. (1966): The General Inquirer: A
Computer Approach to Content Analysis, MIT Press.
18. Tetlock, P.C., Saar-Tsechansky, M., Macskassy, S. (2008): More Than Words: Quantify-
ing Language to Measure Firms’ Fundamentals. The Journal of Finance 63(3), 1437-1467.
19. Mitchell, L., Frank, M.R., Harris, K.D., Dodds, P.S., Danforth, C.M. (2013): The Geogra-
phy of Happiness: Connecting Twitter Sentiment and Expression, Demographics, and Ob-
jective Characteristics of Place. PLoS One 8(5).
20. Remus, R., Quasthoff, U., Heyer, G. (2010): SentiWS – a Publicly Available German-
language Resource for Sentiment Analysis. In: Proceedings of the 7th International Lan-
guage Resources and Evaluation (LREC’10).
21. Võ, M.L-H., Conrad, M., Kuchinke, L., Urton, K., Hofmann, M.J., Jacobs, A.M. (2009):
The Berlin Affective Word List Reloaded (BAWL-R). Behavior Research Methods 41(2),
534-538.
22. Loughran, T., McDonald, B. (2011): When is a Liability Not a Liability? Textual Analysis,
Dictionaries, and 10-Ks. The Journal of Finance 66(1), 35-65.