DHQ: Digital Humanities Quarterly
2022
Volume 16 Number 1
Volume 16 Number 1
A Text Network Analysis of Discursive Changes in German, Austrian and Swiss New Year’s Speeches 2000-2021
Abstract
New Year’s speeches held by leading politicians–mostly prime ministers or presidents–have a long and firm tradition in Europe and have become an institution instead of being a crowing event for conciliatory efforts. From the perspective of political communication, New Year’s speeches fulfil a triple function in the intersection of the past, the present and the future. First, they summarise the past year from the perspective of the political leadership and, hence, recall, reconstruct and remind the most important events of the year. Second, New Year’s speeches describe the present and, thus, can be understood and analysed as reality constructions, as windows to the current state of affairs. And third, New Year’s speeches serve as road maps to the future, into the new year. In this sense, a New Year’s speeches summarises the most important future challenges, expectations and opportunities.
This article stems from the assumption that a nonlinear analysis of textual data based on network analysis could provide us with new ontological understanding about structural coherence and holes within a document corpora. It adopts a different viewpoint to discourse analysis based on social network analysis. The paper introduces a nonlinear way to analyse texts as networks in order to visualise and analyse how concepts are connected and to explore structural closeness and holes within a corpus of unstructured textual documents.
The results indicate a discursive turn in Germany, Austria and Switzerland after the breakout of the global financial crisis in 2008. But at the same time, the results evidence similarities across the three countries how the crisis and its impact was framed to discourses. In all countries, the use of concepts related to crisis and insecurity has increased dramatically since 2008. However, this vocabulary is not solely limited to the financial crisis. The “insecurity and crisis” frame referred both to the financial crisis, the armed conflict in Ukraine and to the refugee crisis.
1 Introduction
During the past decade advances in text mining have demonstrated that political
issues and topics can be identified from unstructured textual documents with good
accuracy. Examples from digital parliamentary studies or computational discourse
analysis have brought up evidence how especially topic modelling can be applied on
large document corpora to gain overviews over thematic structures of political texts,
but also how these structures change over time. Most of these studies analyse large,
even huge textual corpora with thousands of lengthy documents. Such “big text”
analysis is often seen as typical example of how digital methods can – even should –
be applied in social sciences and the Humanities.
In this article I focus on the problem of using digital methods and tools on a
smaller corpus of unstructured textual documents, namely New Year’s addresses, with
varying political content and as addressed to a wide public audience, the citizens in
general. Although politicians often favour longer speeches and verbal battles when it
comes to struggle for political issues, especially the growing relevance of social
media driven political communication has strengthened the role of short textual
documents like tweets, postings, statements or commentaries in political discourse.
Since most short textual contributions are designed to take stand on current affairs
and actual political questions, they offer an interesting empirical material to
explore not only how politicians take stand on current affairs, but also the
structure and dynamics of issue-driven political discourses.
This article is based on the understanding that sort political texts can be a
valuable source when it comes to identifying topics and issues that really matter
under the current political circumstances. Rising a current topic in a short
political text can be seen as an act of politicalisation [Palonen 2007b, 62–66], i.e. an action marking the issue as political by detecting its
political potential and thus expanding the presence of the political and presenting a
particular representation of “reality”
[Koller 2014, 164]. This is especially relevant for addresses and
speeches given by high representatives of a nation like president or prime minister.
All issues taken up in such texts are automatically labelled as being politically
significant, so that the speech itself can be read as an attempt to persuade the
target audience to accept this position [Shim et al. 2015, 56].
The empirical results of this article suggest that New Year’s speeches offer a
valuable and interesting source for the exploration of topical issues enjoying
political and societal relevance. The argument here is rooted in the understanding
that if an issue is taken up in a New Year’s address, the issue is considered as
important and becomes politicised [Palonen 2007a]. Hence, the corpus of
German, Austrian, and Swiss New Year’s speeches offers an interesting source to
explore topical discourses from both national and comparative, cross-national
perspective.
From the methodological perspective the application of digital text mining and
analysis tools on a corpus of short texts consisting of several hundred words in
average requires an adapted approach. This, because most tools for text mining have
been developed for large-to-huge text corpora with millions of words and use
statistical methods to model the corpus. Although these methods produce results also
when applied to small corpora, the results tend to be biased and their reliability
compromised. In order to circumvent these problems I use a rather novel approach of
Text Network Analysis (TNA) using a non-linear approach to re-construct textual
documents as a network of words appearing close to each other in the original
documents [Paranyushkin 2011]. One of the key advantages of TNA is that
since TNA is rooted in (Social) Network Analysis the resulting network can be
visualised and analysed by methods and tools developed for network analysis. Another
advantage is that networks can be created without any arduous preprocessing, although
lemmatisation and stopword removal are strongly encouraged. One main objective of
this article is the introduction of TNA as an alternative, relatively easy to use
explorative tools for both text mining and topic modelling.
The analysis and results presented in this article should evidence the power of TNA
as a topic modelling and exploration tool for short(er) text documents. The analysis
introduces a non-linear way to analyse texts as networks and to present results based
on graphical visualisations making it easier to capture and understand structural
characteristics hidden in the corpus. The results, in turn, provide interesting
insights in thematic continuity and change in political topics in the analysed three
German-speaking countries. Further, a detailed discussion of the results confirm the
accuracy of the main findings of my TNA analysis, thus providing additional support
for the usefulness of the method.
The article is structured as follows. The first main section discusses TNA as an
alternative digital tool for topic modelling. The second section introduces the
empirical corpus used in this article and describes the data preparation process, as
well the analytical steps needed for TNA. The corpus consists of New Year’s addresses
given by political leaders in Germany, Switzerland and Austria between 2000 and 2021.
All speeches are given in German language. The third section presents and discusses
the main findings by reflecting them against contextual factors and previous
research. The article is rounded up with concluding remarks critically summarising
the main findings and achievements.
2 Text Network Analysis
Since the first decade of the 21st century digital research tools and methods have
gained a stronger role and impact among social scientists interested in parliamentary
debates, political speeches or other textual materials on political affairs. A great
majority of such kind of contributions has focused on the identification of the main
content through the analysis of commonly used words, by focusing on the content
different words as used, or through the analysis of co-occurring words.[1]
A known challenge with political speeches is that such texts normally tackle several
topics or issues. However, changes from one topic to another can be be hard to detect
with methods based on a bag-of-words approach. This, because such an approach
transforms the document in a word frequency table. Although such a table is rather
simple to handle and analyse, the other side of the coin is that such a
transformation destroys the original structure of the text, making it impossible to
gain understanding about where different words appear in the original text [Hearst 1994].
[Jelveh et al. 2014] present evidence for an existence of latent shared
ideological positions among experts. In this article I expand this underlying
assumption to include politicians as well, and make the initial assumption that
politicians tend to what I would call “thematic recycling”, i.e. the endeavour
to take up the same – or similar – issues over and over again on different occasions.
Hence, a corpus of political textual documents from a politician can be assumed to
share similar topics across different documents. Further, as in the case of expert
texts we can, with some caution, generalise this assumption and expect to find shared
topics by politicians sharing same or highly similar political insights rooted in
certain ideological positions. Accordingly, we can expect to find shared latent
topics and political positions also in documents delivered by different political
actors sharing a similar political landscape.
Thus far most studies dealing with the problem of identifying topics across documents
in a document collection have applied methods of topic modelling, mostly Latent
Dirichlet Allocation (LDA) [Blei et al. 2003]. There are many good reasons for
this, maybe the most important being its flexibility in defining the “document,”
allowing the researcher to condition the topic distribution on paragraphs or time
chunks, for example. However, as a method based on statistical modelling LDA also has
some important caveats and pitfalls limiting its analytical power especially as
regards shorter text documents and smaller corpora. One limitation being highly
relevant for this article is that most LDA-based applications require datasets of
several hundred documents, at minimum, to ensure validity and reliability of the
results [Light 2014, 117].
Although identifying different topics is an interesting challenge of explorative text
analysis, this article tackles a bit different challenge: how to explore thematic
pathways across documents in a text document corpus. Accordingly, the aim is to
explore, on the one hand, how words are used within statements, and how thematically
similar statements “glue” different texts together across contexts, on the
other. These thematic pathways building on statements and connecting different
contexts should help us to understand the deep thematic backbone underlying the
document corpus [Busse 2013, 43ff].
The analysis presented in the next section is an application of Text Network Analysis
(TNA), a method originally presented by Paranyushkin (2011) and rooted in theoretical
foundations of Social Network Analysis (SNA). According to Paranyushkin TNA seeks to
discover “repetitive patterns derived from the text’s structure,
using their connectivity and the intensity of interactions between them as the
only criteria for their belonging together”
[Paranyushkin 2011, 5]. It is a non-linear approach transforming
texts to a network of words by respecting the original structure of the texts. [Shim et al. 2015, 58], who use TNA to identify policy frames of nuclear
energy policy across different countries, summarise the method by stating that it
seeks “to identify salient words and concepts in order to extract
underlying meanings and frames from the structure of concept networks.” As
the contributions of both Paranyushkin and Shim et al. evidence TNA offers a very
powerful tool with a robust fundament in network analysis and a rather simple and
straightforward application. Further, text networks can be created as flexibly
conditioned on e.g. sliding window techniques, sentences, or paragraphs. This gives a
researcher additional flexibility to test different approaches to figure out the most
suitable text unit for the analysis.
This article stems from the assumption that TNA as a non-linear analysis of textual
data could provide us with new ontological understanding about structural coherence
and holes within a document corpora. This new ontological understanding, however, is
not only expected to shed light on how concepts connect in order to form statements.
Instead, it can also help us to explore structures relevant for the understanding of
how statements are organised within and across contexts. Through the application of
network visualisations and the analysing of structural properties of text networks
the underlying data can be explored from alternative perspectives in order to trace
back discursive patterns within the text corpus.
One key difference of TNA compared to more traditional tools and techniques of text
mining is that TNA understands its input as network data consisting of
nodes (or vertices) connected by edges and
having an own structure (topology) [Paranyushkin 2011]
[Ruohonen 2013]
[Scott 2013]. A common technique is to use word-based units – single
words or n-grams - as nodes and co-occurrences as edges. Although this is the most
common way, nothing actually prevents the researcher from using e.g. phrases as
nodes, just to mention another example illustrating the analytical flexibility of
TNA. Since TNA works based on network theory, it does not require any predefined
ontological structure or pattern of text, nor does it superimpose any extra layer of
ontologies on top of the textual data, which may - as [Paranyushkin 2011, 4] correctly states – create “a strong subjective (and
even cultural) bias into the structure of the resulting text graphs.”
Further, the researcher is not limited to the use of text theories or text mining
tools (although nothing prevents her to use this kind of theories or tools in the
analysis), but can apply a wide range of network analysis tools and methods to
explore, analyse and visualise structural properties of the text networks. In other
words, TNA gives the researcher the possibility to explore and analyse textual units
conditioned to a certain structure and, thus, to explain why different textual units
take different positions and roles by focusing on the network structure of the
text.
3 Data and Workflow
The empirical case study in this article is based on New Year’s speeches of the three
German-speaking countries, Germany, Austria and Switzerland[2]
covering the years from 2000 to 2021. These three countries were selected, first,
based on the shared language providing us the possibility to compare the use of
language of high politics across these countries. Second, all three countries belong
to the same cultural-political space, Central Europe, with a strong, continental
orientation and a centuries-long experience and tradition of German cultural,
economic, and political influence, but also dominance [Jordan 2005].
Besides the two more or less uniting criteria, the third selection criterion was a
dividing factor: Germany and Austria are members in the European Union (EU), whereas
Switzerland is only indirectly involved in European integration through its
membership in the European Free Trade Association (EFTA) and a set of bilateral
agreements with the EU. Switzerland is also member of the Schengen area allowing any
person, irrespective of nationality, to cross the internal borders without being
subjected to border checks. Currently, most of the EU countries plus the non-EU
countries Iceland, Norway, Switzerland and Liechtenstein have joined the Schengen
area. Against this background I expect these three countries to form an “imagined
community” (Benedict Anderson) and to share certain values, interests and world
views reflected by the New Year’s addresses. Further, I expect German and Austrian
New Year’s speeches to share discursive patterns related to EU affairs, especially to
the euro crisis (since 2008), the conflict in Ukraine (since 2014), and the refugee
crisis (since 2015).[3] I also could explore the impact of the global Covid-19 pandemic
disease on both the content and vocabulary across the three countries yet this impact
is visible only in the speeches for the year 2021.
New Year’s speeches held by leading politicians–mostly prime ministers or
presidents–have a long and firm tradition in Europe and “have
become an institution instead of being a crowing event for conciliatory
efforts”
[Portman 2014, 89]. From the perspective of political
communication, New Year’s addresses fulfil a triple function in the intersection of
the past, the present and the future. First, they summarise the past year from the
perspective of the political leadership and, hence, recall, reconstruct and remind
the most important events of the year. Second, New Year’s speeches describe the
present and, thus, can be understood and analysed as reality constructions, as
windows to the current state of affairs. And third, New Year’s speeches serve as road
maps to the future, into the new year. In this sense, a New Year’s speeches
summarises the most important future challenges, expectations and opportunities.
There exist only a small number of dedicated studies on New Year’s speeches [Heikkinen 2006]
[Portman 2014]
[Purhonen and Toikka 2016]. I could not identify any specific reason for this,
despite that New Year’s speeches are dominantly given in non-English speaking
countries in Europe. As a text genre, however, New Year’s addresses are comparable
with the Presidential Inaugural speeches of the U.S. and this article
methodologically benefits from the study of [Light 2014] on the United
States’ Inaugural Addresses. From the perspective of social sciences I am not
primarily interested in the use of language, but issues mentioned and described in
the addresses. This viewpoint is rooted in the assumption that a political
phenomenon, topic or event gains in importance if mentioned in a New Year’s address.
Here I also follow the idea that New Year’s speeches offer insights into shifts in
political priorities, interests, and opinions [Light 2014, 117].
Against this background I am convinced that a longitudinal, computational analysis of
German, Austrian and Swiss New Year’s addresses can help us to a better understanding
of how the political landscape has changed between 2000 and 2021, but also where the
differences and similarities can be found between these three countries.
The speech corpus covers a period of 22 years and contains 64 speeches. There are two
missing years – 2002 and 2017 – in the Austrian speech collection. In Austria Federal
President delivers the New Year’s speeches and between 2000 and 2004 the speeches
were delivered by Thomas Klestil, between 2005 and 2016 by Heinz Fischer, and since
2018 by Alexander Van der Bellen. Although speeches from 2005 onwards can be
downloaded from the official website of
the Austrian Federal President. Fortunately, the speeches of 2000-2001 and 2003-2004
could be found in different Austrian media archives. As regards the year 2002,
neither the President’s office nor the National Archive of Austria could provide any
information about where this particular New Year’s speech could be found, so that I
was forced to drop this year from the corpus. In 2017 Austria had an extraordinary
political situation with no Federal President in office, so that no New Year’s speech
was given.
In Germany, the New Year’s speeches are held by Federal Chancellor and full texts are
available on the official website.[4]. The
first six speeches (2000-2005) were held by Chancellor Gerhard Schröder, the
remaining sixteen from 2006 to 2021 by Chancellor Angela Merkel. In Switzerland, the
annual New Year’s speech is delivered by the President of the Confederation and all
speeches since 1980 are available online on the official website of the President of the Confederation. As the term of the
Swiss President is just one year all New Year’s speeches are held by a different
person (except Micheline Calmy-Rey, Pascal Couchepin and Moritz Leuenberger, all of
them re-elected and with two speeches).
Despite the varieties in the use of language in the German-speaking Europe, formal
speeches by top-level politicians can be seen as an appropriate source of data, as
policy discourses are comprised of policy addresses and speeches about policy issues
[Van Dijk 1977].
Figure 1.
Length (# of terms) of New Year’s speeches in Germany, Austria and Switzerland
2000-2021.New Year’s speeches are traditionally broadcasted over TV or radio within a
relatively short time of 10 to 15 minutes. Consequently, these speeches are relative
compact in size. The average speeches length of a German new New Year’s speech was
892 tokens and standard deviation (sd) was 111. In Austria and Switzerland the
average lengths were 719 (sd=58) and 562 (sd=159) respectively (Figure 1). Quite
understandably the amount of unique tokens is somewhat lower, since certain words are
used several times during a speech. For German New Year’s addresses, there were 362
unique tokens in average (sd=33), for Austrian and Swiss speeches the average number
of unique tokens was 317 (sd=27) and 255 (sd=60) respectively. Swiss New Year’s
speeches seem to have become shorter over time, whereas both German and Austrian New
Year’s speeches remain relative stable over time (Figure 1). Generally speaking a
country’s size seem to matter here, as German New Year’s speeches are slightly longer
than the Austrian or Swiss addresses. My explanation for this observed difference
stems from the differences in political and economic importance. Germany’s central
role in European politics results in broader political agenda, so that there are more
issues and topics to be addressed to in a New Year’s address.
As regards the analysis workflow, all collected speeches were saved in separate files
in plain text format. These files were then imported into RStudio, a graphical end-user environment
for the statistical package R. The
research data was created and analysed in the four steps. First, I used the package
’udpipe’[5] to tokenise,
part-of-speech tag, and lemmatise the documents. After this step, the data was
structured as a data table consisting of 53893 words and 5277 unique lemmata,
enriched with descriptive metadata about the country and the year, as well sentence
indices. In the third step, I used text mining and explorative data analysis tools
from the package ’tidytext’ to carry out different text analyses and to transform textual
data to different sets of text network data. The applied text mining methods will be
described in the next section. For the text network network creation I used both
bi-grams by linking two consecutive words, as well skipgrams of length two by
combining non-adjacent words that were three, four or five words apart.[6] Both the bi-grams and
skipgrams were created within the boundaries of the same sentence based on the
sentence index included in the data structure. (2) in the case of concepts used in a
single country, only concepts which occur more than three times are included. Text
networks were visualised and analysed with ’visone’, a fully-fledged, platform
independent software offering powerful layouts for network visualisations and a
comprehensive set of tools for network manipulation, transformation and analysis.
4 Results
Figure 2.
New Year’s speeches similarity network (2000-2021).Figure 2 presents the addresses network for the whole corpus based on the similarity
of the New Year’s addresses. Similarity is here calculated as cosine measure, a
term-based comparison between documents in a document corpus, and has a value range
from 0 (no similarity at all) to 1 (full similarity).[7] I reduced noise in the document-term matrix
by removing both stopwords and all terms except nouns, verbs, pronouns, and proper
nouns. In the network visualisation nodes represent different New Years speeches and
are coloured with country-specific colours so that German addresses are coloured with
yellow, Austrian with red, and Swiss speeches with white. I have also added labels to
help the reader to identify different addresses. Nodes are connected by edges, of
which width is visualised proportional to the cosine measure of the two speeches
connected by that edge. In my data the cosine measures ranged from 0.103 (between the
Swiss speeches 2006 and 2020) to 0.641 (between the German speeches 2009 and 2010),
but I have removed edges with cosine measure below 0.400 in order to improve the
readability of the network visualisation. Further, I used a force-directed layout and
optimised it to position speeches with higher similarity closer, those with lesser to
each other so that the clusters of more similar speeches are easier to identify.
As regards the similarity network two observations seem appropriate. First, the New
Year’s speeches evidence a stronger similarity within, and weaker similarity across
countries. This is clearly visible in the network structure as well, as speeches from
different countries form clear and distinct cluster in the visualisation (see Figure
2). Overall, the average similarity for cross-country speeches is 0.318 (min=0.114,
max=0.514), for national speeches the average similarity is 0.384 (min=0.103,
max=0.614). If we set the threshold value for cosine measure to 0.5, only two
cross-national combinations, namely Austria 2018 – Switzerland 2020 (0.505) and
Germany 2014 – Switzerland 2018 (0.514), exceed this limit. Varieties of national
political agenda and circumstances seem to explain these observed differences, thus
underlining the primacy of national topical issues. An interesting result is also
that the cosine similarity between German and Austrian speeches is significantly
higher (mean=0.355) than that between German/Austrian and Swiss speeches
(mean=0.277). This difference is at least partly explained by the EU membership, as
certain topical issues both German and Austrian speeches tackle are connected to EU
politics.
And second, similarity seems to be stronger between between chronologically close
speeches and weaker between speeches with greater temporal distance. I tested this
with a simple Pearson’s correlation test and calculated the correlation between the
cosine measure and the temporal distance in years between the New Year’s addresses.
The correlation was -0.085 (p<0.001), thus confirming the hypothesis that the
similarity measure decreases when the temporal distance between compared speeches
increases. This seems logical when reflected against the fact that New Year’s
speeches tackle topical issues at the end of a year. The longer the temporal distance
between two addresses, the more probable there has been significant changes in
political circumstances affecting the vocabulary used in the addresses. Hence, this
observation also gives support to the argument that New Year’s speeches offer an
interesting and valuable perspective on a year’s prevailing political reality.
Figure 3.
The 20 most important words per country in different periods (tf-idf
analysis).Results from the document-level TNA suggested the existence of variance in vocabulary
between the countries and over time. In order to better understand this variance I
applied a specific text mining method, called term frequency, inverse document
frequency (tf-idf) analysis, to adjust the frequency of a word for how
rarely it is used in the collection of New Year’s speeches of each country. A word’s
inverse document frequency (idf) was then applied to explore words that
are not used very much during the whole period from 2000 to 2021. I divided each
national collection of New Year’s speeches into three distinct periods of 2000-2009,
2010-2019, and 2020-2021. Here the idea was to capture changes in the vocabulary over
time within a country, most probably caused by changes in topical issues or political
circumstances. The results of this tf-idf analysis are presented in
Figure 3. In this barplot we can see for each country the most important words in
each distinct period of time. Hence, the Figure 3 highlights temporal changes within
each country, but preserves national differences. In my opinion three central
findings from this analysis are worth being discussed in detail.
First, the connection to topical political issues and circumstances is especially
evident in Germany across the time periods, but also clearly visible in Austrian
speeches of the last two periods. As regards Germany, during the first period
(2000-2009) words like “wiederaufbau” (rebuilding), “betroffen” (upset),
and mitgefühl“” (sympathy, compassion), together with such words like
“naturkatastrophe” (natural disaster), “region” (region), or
“erschüttern” (shock, shake) belonging to the top-30 tf-idf
words in this period, tackle the 2002 European flood. Germany was the hardest hit
country in Europe and the flood resulted in heated political discussions in the eve
of the Federal elections 2002. In Austrian speeches of this period words like
“jahrhudert” (century), “katastrophe” (catastrophe), and
“hilfsbereitschaft” (helpfulness) document the presence of the 2002 flood –
which hit also Austria quite severely – also in Austrian political discourses.
Another issue in this period was the outbreak of the sovereign debt crisis in the
eurozone in 2008/2009. This economic crisis is present in German speeches through
words like “globalisierung” (globalisation), “wirtschaft” (economy),
“arbeitslosigkeit”, (unemployment). Contrary to Germany, the eurozone crisis
is not very present and visible in Austrian speeches in this period of time. A
possible explanation to this difference might be found in the different roles the two
countries played in the early stage of the eurozone crisis. Germany as the strongest
economy in the eurozone played a central, designing role from the very beginning of
the eurozone crisis [Müller-Brandeck-Bocquet 2010]
[Jsnning and Möller 2016]
[Kundnani 2016]
[Bulmer 2018], whereas Austria were not strongly hit by, nor a central
player in this crisis. In addition to that I would also stress the economy-oriented
political culture in Germany. The strong role of economic questions as a central
shaping factor of German domestic and European politics increases the weight of
topical economic issues in public debates. Against this background it is not
surprisingly that the eurozone crisis occupied a more significant space in public
debates in German compared to Austria [Allen 2005]
[Maull 2018].
During the second period (2010-2019) the focus of German New Year’s speeches shifts
from economy to refugee politics. In 2015 Chancellor Merkel decided to keep German
state borders open to refugees coming mainly from the Middle East. Although this
political turn was captured in positive political slogans like
“Willkommenskultur” (culture of welcoming) or – most prominently – “Wir
schaffen das” (we can do it/ we manage it), the fact that words like
“flüchtling” (refugee), “zusammenhalt” (sticking together),
“herausforderung” (challenge), “außengrenze” (outer border), or
“aufnehmen” (receive, welcome) are among the top-20 terms underline the
change the refugee crisis caused in German public and political discourses during
this period of 2010-2019. From the methodological point of view the identification of
terms with a clear connection to the refugee crisis evidence the usefulness of
tf-idf analysis when it comes to explore discursive changes. Contrary
to Germany, the refugee crisis was not central in Austrian New Year’s addresses.
Instead, the national political and economic setting seem to dominate Austrian
addresses. Terms like “marktwirtschaft” (market economy), “lebensqualität”
(quality of life), or “finanz”(finance) tackle economic consequences of the
eurozone crisis. The word “politikverdrossenheit” (jadedness with politics)
refers to domestic political turbulences behind the rise of the right-wing populist
Freedom Party of Austria (FPÖ) in the mid-2010s. Although this rise was partly
supported by the refugee crisis, it was also an expression of disillusionment with
the political elite.
Second, as expected the outbreak of the Covid-19 pandemic disease in Europe in the
first half of 2020 resulted in a clear change, especially as regards the German and
Austrian New Year’s addresses. Words like “pandemie” (pandemic),
“impfung”/“impfstoff” (vaccination), “abstand” (distance), or
“corona” (coronavirus) poke from the results as indicators for this
discursive change in political and public discourse in Germany and Austria. Once
again, the results also confirm the usefulness of tf-idf analysis for
the exploration of discursive changes through changes in vocabulary over time.
And third, the overall results confirm the hypothesis that Germany and Austria are,
what comes to political agenda and circumstances, as well to topical issues, closer
to each other than to Switzerland. It is, however, a somewhat odd observation that
Swiss speeches seem to avoid topical political issues and to concentrate
on the construction of neutral narratives. A possible explanation for this might be
found in the fact that in Switzerland, as opposed to Germany and Austria, the federal
president giving the New Year’s speech is not the head of state, but just the head of
Switzerland’s Federal Council, and only carries out some representative duties.
Hence, the federal president is merely a mediator between the different parts of the
Swiss federal state.
The results presented thus far indicate two main aspects. First, New Year’s addresses
offer a valuable and reliable source to explore topical issues and tackle changes in
political and public discourses. And second, although the national political agenda
and circumstances seem to have a stronger impact on the content of the addresses,
significant European political incidents and events, e.g. the refugee crisis or the
Covid-19 pandemic disease, seem to cause changes across countries. This leads us to
the final analysis, an attempt to explore shared topics. For this analysis I
constructed text networks based on bigrams, i.e. word co-occurrences of two adjacent
words, and of skipgrams, i.e. combinations of two non-adjancent words that are three
to five words apart [Jelveh et al. 2014, 1805]. The bi- and skipgrams
were created for words occurring within the same paragraph. Since the idea is to
analyse topics across countries all word pairs co-occurring in New Year’s speeches of
one single country only were removed. In the last step I calculated a co-occurrence
weight measure by dividing the number of speeches a word pair co-occurs by the total
number of speeches (64) in the multi-national corpus. The interpretation of this
weight measure is rather straightforward: the higher the value, the more significant
the word pair for the cross-country content.
Figure 4.
New Year’s speeches core bigram network across the three countries
(2000-2021).
Figure 5.
New Year’s speeches core skipgram network across the three countries
(2000-2021).Figure 4 visualises word co-occurrences network across countries based on bigrams,
whereas Figure5 visualises word co-occurrence network across countries based on
skipgrams. Hence, both visualised text networks present shared word co-occurrences
central for the multi-country corpus of German, Austrian, and Swiss New Year’s
addresses, i.e. word pairs that can be interpreted as fundamental for the – put in
Benedict Anderson’s (1991) famous terminology – imagined discursive community of
these three German-speaking countries. The bigram-based core text network consists of
336 words and 565 co-occurrences, whereas the skipgram-based core text network
consists of 423 words and 1157 co-occurrences.
In order to visually highlight the most important structural properties both
visualisations apply identical visualisation effects:
- Node size is mapped to the node’s degree centrality value and node label size is mapped the nodes betweenness centrality measure. In network theory centrality, in general, indicates a node’s position in the network and can be calculated either relative to a node’s direct neighbours or the whole network. A node’s degree is the simplest centrality measure and equals to the number of connections the node has to other nodes. Betweenness, as the term itself indicates, defines centrality by analysing where a node is placed within the network. Consequently, a node’s betweenness centrality score is computed by taking into consideration the rest of the network and by looking at how many times a node sits on the shortest path linking two other nodes together, thus helping to identify nodes having “a high probability of occurring on a randomly chosen shortest path between two randomly chosen vertices” [Hasu and Kao 2013] [Prell 2012, 103–104]. Considering meaning circulation across the entire network, the latter capability is assumed to be more relevant, since betweenness centrality “shows the variety of contexts where the word appears, while high degree shows the variety of words next to which the word appears” [Paranyushkin 2011, 13]. This difference is important to keep in mind and therefore I use betweenness centrality to measure a concept’s general status in the text network. As both [Paranyushkin 2011] and [Shim et al. 2015] point out concepts with high betweenness and degree centrality play a meaning circulation role across texts in the document corpus, whereas concepts with low centrality measures are peripheral concepts and, thus, typical only for a certain part of documents. Between these two extremes are located concepts with high betweenness but low degree centrality and concepts with low betweenness but high degree centrality. The former play an important role as bridging concepts between local communities, the latter, in turn, are local hubs within a cluster [Shim et al. 2015, 59f]. This mapping strategy allows us to easily identify the most important and influential words in the graph.
- Node colour is mapped to the topic cluster the node belongs to. One of the proposed key advantages of TNA is bound with the possibility to exploit network community detection techniques.[8] The underlying idea here is that when textual documents are reconstructed as text networks a word’s position in the network is not random but determined relative to the context it appears in. Consequently, words belonging to the same (or similar) context are assumed to cluster across documents, so that we could be able to identify groups of words in the network structure having dense connections within the group, but sparse connections to other parts of the text next. These groups are here interpreted as topics describing the main thematic content of the document corpus. By leaning on the promising results of [Paranyushkin 2011] and [Shim et al. 2015] I applied a modularity-based community detection method called “Louvain method” based on the assumption, that nodes being more densely connected together than with the rest of the network construct a network community [Blondel et al. 2008]. The method identified in total 13 clusters - in my interpretation: topics - in the bigram-based and 12 topics in the skipgram-based text network.
- Edge width is mapped to the weight of the co-occurrence of the word pair. The more often two words co-occur in the document corpus, the wider the connecting edge is visualised in the network graph. Since the graph includes only word pairs shared by at least two countries, edge effects help us to identify word pairs used together frequently across countries.
- Layout used for the visualisation is a radar-like centrality layout, in which a node’s position is determined by its topic cluster and links are arranged according to their weights. The underlying idea of this layout is to gather nodes belonging to the same cluster on the same circumference. Further, the visualisation makes also connections between the clusters – i.e. word pairs, of which words belong to different topic clusters – easier to identify. As regards the meaning circulation across the multi-country text network word pairs connecting two topics can be analysed from the perspective of structural holes. These connections bridging two separate topics by closing a structural hole in a network can make us aware of possibly existing latent similarities between these two topics [Burt 2004].
| Bigram text network core communities | (Total # of clusters: 13, modularity: 0.55) |
| Cluster | Core content words of the topic(a) |
| #1: Solidarity / Sense of community (55 words in total) | bürgerin/bürger, frieden, geschichte, hilfbereitschaft, sicherheit, stabilität, verantwortung, wiederaufbau, zukunft, zusammenhalt |
| #2: Crisis recovery (41) | einsetzen, ereignis, erinnern, freuen, generation, glauben, krise, verbinden, weltkrieg, überwinden |
| #3: Covid-19 era (36) | erfolgreich, familie, friedlich, meistern, pandemie, phase, resignation, schwierig, zeit, zusammenleben |
| #4: Europe (32) | entwicklung, erweiterung, europäisch, hoffnung, krieg, mitgliedstaat, parlament, terror, union, wählen |
| Skipgram text network core communities | (Total # of clusters: 12, modularity: 0.35) |
| Cluster | Core content words of the topic(a) |
| #1: Crisis recovery (62 words in total) | erinnerung, friedlich, krieg, mauer, notwendig, optimismus, pandemie, vergessen, verlust, wirtschaftskrise |
| #2: Faith in the future (51) | aufschwung, bemühen, gestaltung, neugier, selbstvertrauen, technik, wirklichkeit, wissenschaft, zufriedenheit, zuversicht |
| #3: Economy (46) | arbeitsplatz, beitrag, einsatz, erreichen, global, idee, leisten, offen, unternehmen, wirtschaft |
| #4: Solidarity / Sense of community (45) | alltag, bürgerin/bürger, demokratie, entwicklung, familie, frieden, gemeinsam, gemeinschaft, stehen, welt |
Table 1.
Most important topics in the New Year’s speeches across countries.(a) Core content words include ten (10) most influental words relative to the topic
of the cluster (listed in alphabetical order). Bold words indicate top ranking words
both in degree and betweenness centrality, i.e. words being central for meaning
circulation across the text network.
Table 1 summarises the most significant results of the community detection analysis
of both the bigram based and the skipgram based text network by focusing on the four
largest communities found by the Louvain algorithm. As expected there is some
variance across topics between the bigram and skipgram networks. I consider this
variation as normal when the differences in network data creation is taken into
account. Actually, against this background, the two networks should not be considered
as exclusionary but complementary to each other. I have given each community a
descriptive label based on a manual evaluation of the vocabulary linked to the
community.
There are two shared topics. I have labelled the first “Solidarity / Sense of
community” and it is the largest topic in the bigram and fourth largest topic
in the skipgram text network. Both includes words “bürger/bürgerin” (citizen)
and “frieden” (peace). In the bigram network this topic is with respect to
content more focused on helpfulness and security, indicated through words like
“hilfebereitschaft” (helpfulness), “sicherheit” (security),
“stabilität” (stability), “verantwortung” (responsibility), and
“zusammenhalt” (togetherness, solidarity). The skipgram network, in turn, has
in this topic a stronger bias towards democracy and society embodied by words like
“alltag” (everyday life), “demokratie” (democracy), “familie”
(family), and “gemeinschaft” (community).
This topic appears in different contexts over time. In the German New Year’s speech
in 2006 the topic is used to emphasise the inter-linkages between Germany and the
international, global community:
[9]
Deutschland ist keine Insel. Wir stehen in einem internationalen Qualitätswettbewerb, der alle Bereiche unseres Zusammenlebens betrifft: Welcher Nation gelingt es am besten, die schöpferischen Kräfte ihrer Menschen zu wecken? Wie offen ist eine Gesellschaft für Neues? Was bietet sie jungen Familien? In welchem Land gibt es die besten Schulen und Hochschulen? Wie gut gelingt das Miteinander von Einheimischen und Zuwanderern? (Germany 2006)
In the German speech of 2015, the topic is used in the context of the refugee crisis
to foster solidarity among citizens, but also to strengthen confidence and commitment
to the community and its values:
Genauso klar ist: Nur mit offenen Diskussionen und Debatten können wir Lösungen finden, die langfristig Bestand haben und von Mehrheiten getragen werden. Wir sind es, die Bürger und ihre gewählten Repräsentanten, die entwickeln und verteidigen werden, was dieses unser liberales und demokratisches Land so lebenswert und liebenswert macht. Wir sind es, die Lösungen finden werden, die unseren ethischen Normen entsprechen, und den sozialen Zusammenhalt nicht gefährden. Lösungen, die das Wohlergehen der eigenen Bürger berücksichtigen, aber nicht die Not der Flüchtlinge vergessen. (Germany 2015)
The Austrian speech of 2007, in turn, highlights the role of persons in different
positions of trust for the security, stability and confidence of the community:
Lassen Sie mich abschließend die Gelegenheit benutzen, um den vielen Frauen und Männern herzlich zu danken, die in den verschiedensten Funktionen für unser Gemeinwohl, für unsere Sicherheit, für unsere Gesundheit und für den Zusammenhalt unserer Gesellschaft buchstäblich oft Tag und Nacht beruflich, oder auch als freiwillige Helferinnen und Helfer tätig sind. Der Wert dessen, was hier geleistet wird, kann gar nicht hoch genug eingeschätzt werden. (Austria 2007)
And finally, the following two quotes from the Swiss speeches of 2017 and 2021
establish connections between the economic welfare, sovereignty and stability of the
country and the solidarity and togetherness of the society:
Stabil und erfolgreich wollen wir auch in Zukunft sein. Dazu braucht es mehr denn je den Zusammenhalt. Unsere Gesellschaft ist stark, weil wir erprobt sind im Versöhnen von Ansprüchen. (Switzerland 2017)
Wir Schweizerinnen und Schweizer müssen zusammenstehen. Nur so können wir als Land einstehen für die Interessen von uns allen: für unsere Gesundheit und unser wirtschaftliches Wohlergehen, für Frieden und Verbundenheit, für Freiheit und Unabhängigkeit, – kurz: für alles, was uns seit Langem lieb und teuer ist. (Switzerland 2021)
The second shared topic is labelled “Crisis recovery”. The content vocabulary of
this topics indicates that this topic is not limited to the Covid-19 pandemic disease
only. Instead, this topic seems to be used over time to embed a current, topical
crisis into a wider, historical context. Although the clustering algorithm allocates
topical words like “pandemie” (pandemic) or “wirtschaftskrise” (financial
crisis) to this cluster, in my interpretation other content words like
“generation” (generation), “krieg / weltkrieg” (war / world war),
“überwinden” (overcome), “optimismus” (optimism), “mauer” (wall),
and “erinnern” (remember) stand for this wider historical context.
Both in the Swiss speech in 2010 and German speech in 2011 this topic was used to
frame the eurozone crisis. Both quotations manifest a clear optimism that the crisis
has been successfully fought and the country will recover rather quickly:
Wir verfügen über eine starke Wirtschaft und eine solide Finanzpolitik. Wir haben es geschafft, trotz grosser Wirtschaftskrise die Schulden nicht übermässig wachsen zu lassen. Das wird uns im Aufschwung stärken. (Switzerland 2010)
Deutschland hat die Krise wie kaum ein anderes Land gemeistert. Was wir uns vorgenommen hatten, das haben wir auch geschafft: Wir sind sogar gestärkt aus der Krise herausgekommen. Und das ist vor allem Ihr Verdienst, liebe Mitbürgerinnen und Mitbürger. (Germany 2011)
In 2020, the Austrian speech used this topic in a totally different context combining
the most important crises and challenges of the current time, i.e. climate change,
digitalisation, and migration, as well questions related to gender equality and
welfare state reforms.
Zur Klimakrise kommen weitere große Herausforderungen: Wie werden wir künftig arbeiten? Welche Antworten geben wir in Österreich auf die Digitalisierung? Wie soll sich unser Wirtschaftsstandort entwickeln? Wie gehen wir mit Migration um? Und was tun wir, um Frauenrechte zu stärken? Haben wir ausreichend drüber nachgedacht, das Nötige im Bildungsbereich anzugehen? Welche Reformen sind im Gesundheits- und Sozialsystem notwendig, um soziale Sicherheit und sozialen Zusammenhalt für die Zukunft zu gewährleisten? (Austria 2020)
Both text networks produce also distinct topics. In the bigram network we can
identify the unique topics labelled by myself “Covid-19 era” and “Europe”,
in the skipgram network such topics include “Faith in the future” and
“Economy”. As regards the bigram-specific topics, “Covid-19 era” not
only embodies the strenuousness of the pandemic disease by content words like
“pandemie” (pandemic), “resignation” (resignation), or the word
combination “schwierig” + “zeit” (hard times), but also seeks to create
hope through words like “erfolgreich” (successful), “familie” (family),
“meistern” (control), “zusammenleben” (life together). The topic
“Europe”, in turn, tackles both current European issues like terrorism
(“terror”), European parliamentary elections (“wählen” (vote),
“parlament” (parliament)) and European integration as a wider context
(“europäisch” + “union” (EU), “erweiterung” (enlargement),
“mitgliedstaat” (member state)). This latter topic clearly evidences the role
and status of the EU as the most important political, economic, and geographical
context.
The topic “Covid-19 era” is a mixture of resignation and hope. The Austrian
speech in 2021 well exemplifies this mixture:
Ein neues Jahr liegt vor uns. Wir spüren noch die Last des alten, die Last der Pandemie. Aber viele von uns spüren trotz allem eine hoffnungsfrohe Erwartung, wie sie nur am Beginn von etwas Neuem stehen kann, wenn alle Möglichkeiten offen und alle Träume noch frisch sind. (Austria 2021)
The German address, in turn, acknowledges the dramatic changes caused by the Covid-19
pandemic disease. At the same, however, the quotation has a positive trait, stressing
the supportive and stabilising role of the federal state in the times of crisis:
Die Aufgaben, vor die die Pandemie uns stellt, bleiben gewaltig. Bei vielen Gewerbetreibenden, Arbeitnehmern, Solo-Selbstständigen und Künstlern herrschen Unsicherheit, ja Existenzangst. Die Bundesregierung hat sie in dieser ganz unverschuldeten Notlage nicht allein gelassen. Staatliche Unterstützung in nie dagewesener Höhe hilft. Verbesserte Kurzarbeitsregeln greifen. Arbeitsplätze können so bewahrt werden. (Germany 2021)
Contrary to Austria and Germany, the Swiss speech of 2021 is marked by negative
ambience, even resignation underlining the deep societal impact the Covid-19 pandemic
disease has caused in Switzerland:
Selten haben wir Vergleichbares erlebt: Unsere Tätigkeiten kamen zum Stillstand. Die ganze Gesellschaft befand sich in noch nie dagewesener Isolation. Wir mussten lernen, ohne Händeschütteln auszukommen. Dieses wichtige Begrüssungsritual gefährdete plötzlich unsere Gesundheit. (Switzerland 2021)
The topic “Europe” is mostly used, as the following quotations exemplify, to
remind the citizens about the fundamental importance of European integration as a
uniting community of Europeans, but also as provider for economic and political
security in a global world:
Es geht nun darum, dass die Union nicht nur wirtschaftlich, sondern auch verstärkt politisch und emotional zusammenwächst. Die Bürger müssen spüren, dass es sich lohnt, in diesem erweiterten Europa zu leben und zu arbeiten. (Austria 2003)
In den vergangenen Jahren habe ich oft gesagt, dass es auch Deutschland auf Dauer nur dann gut geht, wenn es auch Europa gut geht. Denn nur in der Gemeinschaft der Europäischen Union können wir unsere Werte und Interessen behaupten und Frieden, Freiheit und Wohlstand sichern. (Germany 2020)
Wirtschaftlich sind wir weltweit verflochten. Geografisch, historisch und kulturell sind wir ein europäisches Land. Die Beziehungen zu unseren Nachbarländern und zu den anderen Mitgliedstaaten der Europäischen Union regeln wir in bewährter Weise bilateral. (Switzerland 2005)
Within the skipgram network, the topic “Faith in the future” represents a strong
faith in technological and scientific development as the most important fundament for
a better future. This is especially evident when we consider such content words like
“technik” (technology), “wissenschaft” (science), “aufschwung”
(boom, economic expansion), “selbstvertrauen” (self-confidence), and
“zuversicht” (confidence). Interestingly, this topic also contains words with
strong, dynamic connotation to the shaping the future (“gestaltung”) and the
putting oneself out (“bemühen”). How this topic is used to generate optimism
towards the future is well illustrated by the following quotations:
Wie wichtig es ist, auch im scheinbar größten Durcheinander Gelassenheit, Mut und Zuversicht zu bewahren. (Austria 2020)
Zusammenhalt, Offenheit, unsere Demokratie und eine starke Wirtschaft, die dem Wohl aller dient: Das ist es, was mich für unsere Zukunft hier in Deutschland auch am Ende eines schweren Jahres zuversichtlich sein lässt. (Germany 2017)
Dennoch bin ich zuversichtlich. Denn ich glaube an die Fähigkeit des Menschen zu gestalten. Es gibt immer einen Weg. Manchmal braucht es Mut und Kraft. (Switzerland 2010)
Finally, the topic “Economy” is characterised by very concrete, economic content
words like “arbeitsplatz” (job, post), “global” (global[isation]),
“unternehmen” (company), and “wirtschaft” (economy). But this topic also
has a forward-looking content embodied by words like “erreichen” (reach),
“idee” (idea), or “leisten” (perform). An excellent example of how this
topic is present in the speeches is the following quotation from the German speech of
2008: “Deutschland kann seine alte Kraft als das Land der Sozialen Marktwirtschaft
wieder neu unter Beweis stellen, der Verbindung von Freiheit und Gerechtigkeit,
Fleiß und Unternehmergeist”. The same spirit can be found in the following
Austrian and Swiss addresses:
Unsere Bevölkerung hat nicht nur ein Recht auf sichere Grenzen und innere Sicherheit, sie hat auch ein Recht auf ein möglichst hohes Maß an wirtschaftlicher Stabilität und sozialer Sicherheit. (Austria 2013)
Starke Unternehmen sind der beste Garant für Arbeitsplätze. Und Arbeitsplätze sind die Basis für soziale Sicherheit und Wohlstand. Noch gehört die Schweiz zu den wirtschaftlich erfolgreichsten Ländern. Sorgen wir mit freiheitlichen Rahmenbedingungen dafür, dass das so bleibt. (Switzerland 2016)
Overall, the topics identified by the Louvain method seem appropriate and reliable in
the context of these three countries. The main topics discussed above not only
connect well to the political and economic reality in these countries or in Europe in
more general terms, but also evidence that TNA tools offer – as the previous studies
of [Paranyushkin 2011], [Light 2014], or [Shim et al. 2015] already suggest – a reliable alternative to traditional
methods of text mining. Further, the differences between topics identified in the
bigram and the skipgram based network seem to give support to the study of [Jelveh et al. 2014] that the complementary use of both methods can help us to
reliably identify topics that use similar vocabulary but show differences in the text
structure. And finally, it is worth being noted that the most important content words
of each topic contain only a few words with both a high degree centrality and a high
betweenness centrality. This means that the most important content words are those
enjoying higher relevance within the topic and, thus, are descriptive for the topic.
A closer look at words used for meaning circulation across topics and countries
reveals that these words include words related to peace (“frieden”,
“friedlich”) and war (“krieg”), security (“sicherheit”), crisis
(“krise”), community (“gemeinschaft”), economy (“wirtschaft”), and
future (“zukunft”). All these “glueing” words are less topical and more
contextual, thus framing and embedding the topical issues into a wider context of
European (integration) history.
5 Discussion
This article has sought to exemplify the utility of text network analysis as a
computational method for exploratory text and content analysis. Further, in the
leaning on similar previous studies, this article also understands TNA as an
promising and powerful alternative to traditional methods of topic modelling,
especially to LDA. Against the background of these objectives and based on the
results of the empirical analysis the following three conclusions are worth being
discussed.
First, one major objective of this article was to present and evidence the usefulness
of computational, exploratory text analysis for gaining new insights of the structure
and dynamics of a collection of short textual documents. In this respect, the results
are promising. Especially I would like to highlight the possibilities of using
network clustering methods for the exploration of topics hidden in the corpus. As the
result indicate, network clustering algorithms can identify contextually meaningful
and relevant clusters. As I have shown the applied approach can help the researcher
to identify appropriate text passages for different topics, serving as a starting
point for an in-depth context analysis.
Second, results from the empirical analysis evidence the analytical power of the two
computational methods applied. On the one hand, the rather traditional
tf-idf method helped to gain insights and a better understanding of
changes over time and of differences between the three countries. By increasing the
weight of words that are more common in a certain period I could tackle changes in
the content of the speeches over time. On the other hand, TNA proved to be a rather
simple, yet powerful alternative to traditional topic modelling methods. The
validation of the results of the community detection against samples from the
original sources showed that the identified clusters not only had a meaningful
content relevant for all countries, but they also were, according to their modularity
values, valid and reliable. The two methods, when applied in a complementary manner,
were very useful in identifying changes in the content over time and across nations,
but also in exploring topics binding these three countries together. Here we could
also identify one clear difference between the three countries. For Germany and
Austria Europe seem to play an important role and many speeches reflected the
fundamental importance of Europe, especially the European Union, as the
political, economic, and historical context for both countries. Swiss New Year’s
speeches remain somewhat abstract and globally oriented in this respect, thus
reflecting the long tradition of the neutral Swiss foreign policy [Gabriel and Fischer 2003]
[Goetschel and Schwarz 2005]
Third, although the results are encouraging both in empirical and methodological
sense, the limitations of text network analysis need to be addressed. As Diesner
et al. (2012) points out, validation of the results can be difficult for densely
connected large-scale networks. Further, techniques for text preprocessing, node
identification, and link construction must be decided before mining network structure
from text data, since these decisions “could strongly influence
the structure of resulting networks”
[Shim et al. 2015, 75]. As a comparison of the bigram and skipgram
networks evidenced, different methods of network data creation may produce different
results, so that data preprocessing and network analysis techniques should be
selected with care and be closely aligned with research questions and objectives.
Overall, the analysis presented in this article is well in line with these previous
works and supportive for the idea that text network analysis could offer an
interesting alternative method for computational content analysis. Since the method
seems to work quite reliable also with smaller data sets I can only encourage
colleagues interested in this kind of analysis to test TNA tools.
Notes
[1] For a
good overview of different methods for content mining, see e.g. [Aijmer and Stenström 2004]
[Lee and Kim 2010]
[Stuart and Botella 2009]
[Weiss et al. 2010]
[Diesner and Tambayong 2012]
[Leetaru 2012]
[Ignatow and Mihalcea 2016]
[Jacobs and Tschötschel 2019]
[2] Switzerland is in
fact a multilingual country with four official languages, German included.
[3] For a similar design, see [Shim et al. 2015] who
apply text network analysis on nuclear policy documents from six countries in
order to analyse how the Fukushima accident affected nuclear energy policy
frames.
[4] It
should be noted that in Germany also the Federal President (Bundespräsident) gives a speech on Christmas. This Christmas speech
is, what comes to its function as a political speech act, very similar to the New
Year’s speech. Since the purpose of this article is to compare New Year’s
addresses, German Christmas speeches are not included in this study.
[5] The package is available and documented at:https://bnosac.github.io/udpipe/docs/doc2.html. For a technical and
scientific description of the package, see Straková 2017.
[6] This
technique was adopted from Jelveh et al. (2014).
[7] For details, see [Leydesdorff 2005]
[8] For a summary, see [Fortunato 2010]
[Wu et al. 2013]
[9] References to speeches can be located at the following: Austrian New Years
speeches 2000-2021. Available at the website of the Federal President of
Austria, German New Years speeches 2000-2021. Available at the website of
the German Chancellor's Office, Swiss New Years speeches 2000-2021. Available
at the website of the Swiss Federal Council
Works Cited
Aijmer and Stenström 2004 Aijmer, K. and A.-B.
Stenström (Eds.) (2004). Discourse Patterns in Spoken and
Written Corpora. Amsterdam: John Benjamins Publishing.
Allen 2005 Allen, C. S. (2005). “Ideas, Institutions and Organized Captitalism: Germany, Europe and 21st Century
Economic Policy Models.”
Conference Papers – American Political Science
Association, 1–25.
Anderson 1991 Anderson, B. (1991). Imagined communities: reflections on the origin and spread of
nationalism (Rev. and extended ed. ed.). London: Verso.
Blei et al. 2003 Blei, D. M., A. Y. Ng, and M. I. Jordan
(2003). “Latent Dirichlet allocation.”
Journal of machine learning research 3(4-5), 993–1022.
Blondel et al. 2008 Blondel, V. D., J.-L. Guillaume,
R. Lambiotte, and E. Lefebvre (2008). “Fast unfolding of
communities in large networks.”
Journal of Statistical Mechanics: Theory and
Experiment (P10008).
Bulmer 2018 Bulmer, S. (2018). “Germany and the european union: Post-brexit hegemon?”
Insight Turkey 20(3).
Burt 2004 Burt, R. (2004). “Structural Holes and Good Ideas.”
American Journal of Sociology 110(2), 349–399.
Busse 2013 Busse, D. (2013). “Linguistische Diskurssemantik: Rückschau und Erläuterung nach 30 Jahren.”
In D. Busse and W. Teubert (Eds.), Linguistische Diskursanalyse:
neue Perspektiven, pp. 31–53. Wiesbaden: Springer VS.
Diesner and Tambayong 2012 Diesner, J., K. M. Carley,
and L. Tambayong (2012). “Extracting socio-cultural networks of
the Sudan from open-source, large-scale text data.”
Computational and Mathematical Organization
Theory 12(3), 328–339.
Fortunato 2010 Fortunato, S. (2010). “Community detection in graphs.”
Physics Reports (486), 75–174.
Gabriel and Fischer 2003 Gabriel, J. M. and
T. Fischer (Eds.) (2003). Swiss Foreign Policy,
1945-2002. Houndmills: Palgrave Macmillan.
Goetschel and Schwarz 2005 Goetschel, L.,
M. Bernath, and D. Schwarz (2005). Swiss Foreign Policy:
Foundations and Possibilities. London: Routledge.
Hasu and Kao 2013 Hsu, Y.-Y. and H.-Y. Kao (2013). “CoIN: a network analysis for document triage.”
Database, 1–11.
Hearst 1994 Hearst, M. A. (1994). “Multi-Paragraph Segmentation of Expository Text.”
Proceedings of the 32nd Annual Meeting of the Association for
Computational Linguistics (ACL 1994).
Heikkinen 2006 Heikkinen, V. (2006). “Uudenvuodenpuheiden piirteitä 1935–2006 ja näkymiä vallan
medioitumiseen.” In T. Nikko and P. Pälli (Eds.), Kieli ja teknologia. Talous ja kieli IV, Number B76 in Publications of
the Helsinki School of Economics, pp. 173–194. Helsinki: Helsinki School of
Economy.
Ignatow and Mihalcea 2016 Ignatow, G. and R. F.
Mihalcea (2016). Text Mining: A Guidebook for the Social
Sciences. London: SAGE.
Jacobs and Tschötschel 2019 Jacobs, T. and
R. Tschötschel (2019). “Topic models meet discourse analysis: a
quantitative tool for a qualitative approach.”
International Journal of Social Research
Methodology 22(5), 469 – 485.
Jelveh et al. 2014 Jelveh, Z., B. Kogut, and S. Naidu
(2014). “Detecting Latent Ideology in Expert Text: Evidence From
Academic Papers in Economics.” In Proceedings of the
2014 Conference on Empirical Methods in Natural Language Processing
(EMNLP), pp. 1804–1809.
Jordan 2005 Jordan, P. (2005). “Großgliederung Europas nach kulturräumlichen Kriterien.”
Europa Regional 13(4),
162–173.
Jsnning and Möller 2016 Janning, J. and A. Möller
(2016). “Leading from the centre: Germany’s new role in
Europe.” ECFR Policy Brief, European Council on Foreign Relations.
Koller 2014 Koller, V. (2014). “From text to political positions via discourse analysis: Introduction.” In
B. Kaal, I. Maks, and A. van Elfrinkhof (Eds.), From Text to
Political Positions. Text Analysis across Disciplines, pp. 163–169.
Amsterdam: John Benjamins Publishing.
Kundnani 2016 Kundnani, H. (2016). German Power. Das Paradox der deutschen Stärke. München:
C.H.Beck.
Lee and Kim 2010 Lee, S., J. Song, and Y. Kim (2010).
“An Empirical Comparison of Four Text Mining Methods.”
Journal of Computer Information Systems 51(1), 1–10.
Leetaru 2012 Leetaru, K. H. (2012). Data mining methods for the content analysts. New York:
Routledge.
Leydesdorff 2005 Leydesdorff, L. (2005). “Similarity measures, author cocitation analysis, and information
theory.”
Journal of the American Society for Information Science &
Technology 56(7), 769–772.
Light 2014 Light, R. (2014). “From
words to networks and back: Digital text, computational social science, and the
case of presidential inaugural addresses.”
Social Currents 1(2),
111–129.
Maull 2018 Maull, H. W. (2018). Reflective, Hegemonic,
Geo-economic, Civilian...? The Puzzle of German Power. German
Politics 27(4), 460–478.
Müller-Brandeck-Bocquet 2010 Müller-Brandeck-Bocquet, G. (2010). “Deutsche Europapolitik
unter Angela Merkel: Enge Gestaltungsspielräume in Krisenzeiten.” In
G. Müller-Brandeck-Bocquet, C. Schukraft, N. Leuchtweis, and U. Kessler (Eds.),
Deutsche Europapolitik: Von Adenauer bis Merkel, pp.
253–349. Wiesbaden: VS Verlag für Sozialwissenschaften. 2., aktualisierte und
erweiterte Auflage.
Palonen 2007a Palonen, K. (2007a). Four Times of Politics: Policy, Polity, Politicking and
Politicization, In: K. Palonen (2007b), pp. 55–70.
Palonen 2007b Palonen, K. (2007b). Re-thinking Politics. Essays from a Quarter-century.
Jyväskylä: University Printing House.
Paranyushkin 2011 Paranyushkin, D. (2011).
“Identifying the Pathways for Meaning Circulation using Text
Network Analysis.” Research report, Nodus Labs.
Portman 2014 Portman, A. (2014). From Might to Mandata: Values in Rulers’ Public Speeches in Finland
1809-2000. Publications of the Department of Social Research of the
University of Helsinki 2014:4. Helsinki: Unigrafia.
Prell 2012 Prell, C. (2012). Social Network Analysis: History, theory and methodology. London:
SAGE.
Purhonen and Toikka 2016 Purhonen, S. and A. Toikka
(2016). “'Big Data' haaste ja uudet laskennalliset
tekstiaineistojen analyysimenetelmät. Esimerkkitapauksena aihemallianalyysi
tasavallan presidentin uudenvuodenpuheista 1935-2015.”
Sosiologia 53(1),
6–27.
Ruohonen 2013 Ruohonen, K. (2013). “Graph theory.” Technical report, Tampere University of
Technology.
Scott 2013 Scott, J. P. (2013). Social Network Analysis (3 ed.). London: Sage Publishing.
Shim et al. 2015 Shim, J., C. Park, and M. Wilding
(2015). Identifying policy frames through semantic network
analysis: an examination of nuclear energy policy across six countries.
Policy Sciences 48(1), 51–83.
Straková 2017 Straková, M. S. A. (2017). “Tokenizing, POS Tagging, Lemmatizing and Parsing UD 2.0 with
UDPipe.” Technical report, Charles University, Faculty of Mathematics and
Physics, Institute of Formal and Applied Linguistics, Prague.
Stuart and Botella 2009 Stuart, K. and A. Botella
(2009). “Corpus Linguistics, Network Analysis and Co-occurrence
Matrices. ”International Journal of English
Studies (S1), 1–28.
Van Dijk 1977 van Dijk, T. A. (1977). Text and context. London: Longman.
Weiss et al. 2010 Weiss, S. M., N. Indurkhya, and
T. Zhang (2010). Fundamentals of Predictive Text Mining.
London: Springer.
Wu et al. 2013 Wu, Z., J. Cao, J. Wu, Y. Wang, and C. Liu
(2013). “Detecting Genuine Communities from Large-Scale Social
Networks: A Pattern-Based Method.”
The Computer Journal, 1–15.
URL: http://www.digitalhumanities.org/dhq/vol/16/1/000598/000598.html
Comments: dhqinfo@digitalhumanities.org
Published by: The Alliance of Digital Humanities Organizations and The Association for Computers and the Humanities
Affiliated with: Digital Scholarship in the Humanities
DHQ has been made possible in part by the National Endowment for the Humanities.
Copyright © 2005 -
Unless otherwise noted, the DHQ web site and all DHQ published content are published under a Creative Commons Attribution-NoDerivatives 4.0 International License. Individual articles may carry a more permissive license, as described in the footer for the individual article, and in the article’s metadata.
Comments: dhqinfo@digitalhumanities.org
Published by: The Alliance of Digital Humanities Organizations and The Association for Computers and the Humanities
Affiliated with: Digital Scholarship in the Humanities
DHQ has been made possible in part by the National Endowment for the Humanities.
Copyright © 2005 -
Unless otherwise noted, the DHQ web site and all DHQ published content are published under a Creative Commons Attribution-NoDerivatives 4.0 International License. Individual articles may carry a more permissive license, as described in the footer for the individual article, and in the article’s metadata.
