Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=rdij20
Download by: [Masarykova Univerzita v Brne] Date: 18 September 2017, At: 11:09
Digital Journalism
ISSN: 2167-0811 (Print) 2167-082X (Online) Journal homepage: http://www.tandfonline.com/loi/rdij20
Quantitative analysis of large amounts of
journalistic texts using topic modelling
Carina Jacobi, Wouter van Atteveldt & Kasper Welbers
To cite this article: Carina Jacobi, Wouter van Atteveldt & Kasper Welbers (2016) Quantitative
analysis of large amounts of journalistic texts using topic modelling, Digital Journalism, 4:1, 89-106,
DOI: 10.1080/21670811.2015.1093271
To link to this article: http://dx.doi.org/10.1080/21670811.2015.1093271
Published online: 13 Oct 2015.
Submit your article to this journal
Article views: 383
View related articles
View Crossmark data
Citing articles: 8 View citing articles
QUANTITATIVE ANALYSIS OF LARGE
AMOUNTS OF JOURNALISTIC TEXTS
USING TOPIC MODELLING
Carina Jacobi, Wouter van Atteveldt and Kasper Welbers
The huge collections of news content which have become available through digital technologies
both enable and warrant scientific inquiry, challenging journalism scholars to analyse
unprecedented amounts of texts. We propose Latent Dirichlet Allocation (LDA) topic modelling
as a tool to face this challenge. LDA is a cutting edge technique for content analysis, designed
to automatically organize large archives of documents based on latent topics, measured as
patterns of word (co-)occurrence. We explain how this technique works, how different choices
by the researcher affect the results and how the results can be meaningfully interpreted. To
demonstrate its usefulness for journalism research, we conducted a case study of the New
York Times coverage of nuclear technology from 1945 to the present, partially replicating a
study by Gamson and Modigliani. This shows that LDA is a useful tool for analysing trends
and patterns in news content in large digital news archives relatively quickly.
KEYWORDS automatic content analysis; journalism; nuclear energy; topic models
Introduction: Latent Dirichlet Allocation, Topics and Issues
The shift of news media towards online publication and archiving provides journalism
scholars with new opportunities for studying journalism. At the same time, understanding
the complicated dynamics of this contemporary media landscape requires an
ever-larger scale of analysis, with more outlets and more content per outlet. In this article,
we show how topic modelling, a relatively new method developed in the computational
linguistics field, can help analyse large amounts of text without requiring manual
coding, thus reducing the time and costs of such projects. For a general overview of the
available methods and pitfalls for topic models, see Grimmer and Stewart (2013). In this
paper, we focus on one type of topic model, Latent Dirichlet Allocation (LDA; Blei, Ng,
and Jordan 2003), and demonstrate its use for journalism research. Even though topic
modelling is a promising method for text analysis, with the seminal paper in computational
linguistics (Blei, Ng, and Jordan 2003) published around a decade ago, it is just
starting to be used in the social sciences. Political scientists, who, like journalism
researchers, have both the challenge and the opportunity of newly available online
archives of textual data (such as political speeches, legislative documents and social
media) have started to use topic models to automatically classify these documents.
Digital Journalism, 2016
Vol. 4, No. 1, 89–106, http://dx.doi.org/10.1080/21670811.2015.1093271
Ó 2015 Taylor & Francis
This article was originally published with errors. This version has been
corrected. Please see Corrigendum (http://dx.doi.org/10.1080/21670811.
2015.1111968).
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
Notably, Quinn et al. (2010) classify speeches in the US Senate into topics using topic
modelling. Lucas et al. (2015) apply topic modelling to different types of documents
such as fatwas and social media posts in order to facilitate comparison between countries.
Following such studies, we explore whether topic modelling can be used to classify
journalistic documents into categories that are meaningful to journalism researchers.
LDA, like other topic modelling algorithms, is an unsupervised technique that
automatically creates topics based on patterns of (co-)occurrence of words in the documents
that are analysed. Journalistic texts are thus “coded automatically” for topics,
although it is up to the researcher to interpret the results of the model and to set up
the analysis in such a way that the results are useful for the study at hand. Thus, the
usefulness of the technique for studying journalism crucially depends on the correspondence
between topics and the constructs of theoretical interest. The goal of this article
is to introduce LDA to journalism scholars and to provide a practical guide in applying
the technique to their own research. Concretely, we will deal with three broad topics:
• What is topic modelling? The first part of this article will give a brief and mostly
non-technical description of LDA.
• How to set up an LDA topic model: Secondly, we will describe the different parameters
of the LDA topic model, and discuss issues of validity.
• Theoretical interpretation: The last and most important part of the article discusses
how LDA topic models relate to theoretical constructs of interest to journalism
researchers, especially issues and frames. Using the example of the news discourse
on nuclear technology from 1945 to now, we show how LDA topics mostly
correspond to the important issues in this discourse, comparing our results to the
earlier study by Gamson and Modigliani (1989).
What is Topic Modelling?
Topic models are computer algorithms that identify latent patterns of word
occurrence using the distribution of words in a collection of documents. The output is
a set of topics consisting of clusters of words that co-occur in these documents according
to certain patterns. In an LDA model, each document may contain multiple topics.
Each of the topics has an internal consistency—the words in that topic often occur
together in the documents, and/or do not appear much outside that topic. The
researcher determines what this consistency refers to, and thus how the topic can be
interpreted (Chang et al. 2009). It is this interpretability that determines whether topic
models in general, and LDA specifically, are of use to social scientists.
In the case of journalism research, the collection of topics inferred by the model
would ideally resemble a categorization of issues or frames based on substantive theory,
for example the issue list used by the Comparative Agendas Project that uses categories
such as Macro-economics, Foreign Affairs and Crime to categorize legislative and
journalistic texts (Jones and Baumgartner 2005). However, topics are created by the
LDA algorithm based on patterns of word co-occurrence in documents, which do not
necessarily match theoretical concepts. It seems plausible to equate a topic with the
theoretical concept “issue” or “theme”, but topics could potentially also represent writing
or speaking styles (e.g. words referring to emotions), events (e.g. a natural disaster)
90 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
or frames (e.g. immigrants framed as criminals), which, at least theoretically, are also
formed through co-occurrence patterns of specific words. For example, in the study of
Quinn et al. (2010) of topics in Congressional Record speeches, words such as violence,
drug trafficking, police and prison were interpreted as being the topic Law and Crime I:
Violence/Drugs. Another topic, Abortion, contained words such as baby, abortion, procedure
and life (Quinn et al. 2010, 217). This study thus interprets topics as issues,
whereas DiMaggio, Nag, and Blei (2013, 590–591) refer to “voices” or “frames” when
interpreting different topics. However, what exactly topics represent, and if they
represent different concepts given different input parameters in the model, is ultimately
an empirical question.
To illustrate what topics look like and the interpretation challenges they present,
consider the example newspaper article in Figure 1. This article, which appeared after
the Chernobyl disaster, criticized the Soviet Union for suppressing news about the
event. In the text, each word is highlighted to indicate which topic it was drawn from.
These topics were found automatically by the algorithm, but we interpreted and named
the topics in a way similar to how one would interpret a factor analysis outcome.1
As would be expected given the content of the article, many words are drawn
from a topic that we interpret as the Nuclear Accidents topic, indicated in dark grey
with white letters. This topic includes words such as “Chernobyl”, “disaster” and “radiation”,
but interestingly also contains the words “last Friday”, probably due to the episodic
nature of accident reporting. The other main topic, Cold War, is indicated in
medium grey and contains words such as “Soviet Union” and “Gorbachev” but also
“confidence” and “pledge”. Finally, a number of terms such as “secrecy” and “peaceful
use” of “energy” are drawn from the Nuclear Research topic, and two words are drawn
from other topics.
This example highlights a number of interesting points about LDA. First, this document
is split between two main topics, Cold War and Nuclear Accidents. In a coding
scheme forced to have a single topic per document, it would be very difficult to
choose the “dominant” topic for this article, so in our opinion this accurately reflects
the nature of the article. Second, you can see that not all words are included in the
analysis: most of the words in the article are not used by the algorithm (the text without
highlighting), either because they are non-substantive words such as determiners
or prepositions (“the” or “it”), which we excluded, or because they are too rare (“Elbe”,
“perish”) or too common (“have” but in this context also “nuclear”, since that was used
FIGURE 1
Example article with words from different topics highlighted in the text
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 91
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
to select the articles). Finally, it should be remembered that no a priori coding scheme
was used by the computer, so the topics in this document were found completely
automatically. The fact that the resulting topics (such as Cold War and Nuclear Accidents)
make sense substantively and match the way we would define the subject or
topic of articles—such as the example here—shows that to some extent our notion of
“topic” matches the latent classes or co-occurrence patterns identified by the LDA
algorithm.
Topics such as Cold War and Proliferation can be interpreted as issues, but also as
ways of framing nuclear weapons. In such cases, a particular perspective on an issue
can be signified by specific keywords, which can be seen as framing devices (Matthes
and Kohring 2008). In general, if framing devices correspond to specific (latent) patterns
of vocabulary use, LDA can capture these classes in specific topics, and as such LDA
results can also include the frames used in a corpus of texts.
How to Set Up an LDA Topic Model
Before exploring a case study where we interpret LDA topics, we will explain how
to perform an analysis using this technique. As previously mentioned, LDA processes a
collection of documents and clusters the words in these documents into topics in an
unsupervised way. As such, it is important to start by considering which documents
should be included, and to make sure these documents consist only of the content of
newspaper articles—other textual information such as author name or reader
comments should be removed.
We will discuss the steps we took when conducting the analysis for our case
study below. Our data consisted of 51,528 news stories (headline and lead) from the
New York Times that mentioned nuclear power, published between 1945 and 2013. We
retrieved all news stories from the New York Times online archive (http://developer.ny
times.com) that contained the search terms “nuclear”, “atom” or “atomic” in the
headline or lead.
Preprocessing: Tokenization and Lemmatization
A topic model does not analyse documents directly, but uses a so-called document–term
matrix based on these documents. This matrix lists the frequency for each
term (word) in each document. The first step in creating this matrix is tokenization,
which means splitting the text into a list of words. Although this can be done by splitting
on white space and punctuation, there are good word boundary detection tools
that recognize acronyms, contractions, etc.
It is often better to begin by reducing the size of the matrix by preprocessing
and feature selection, reducing computing time and improving results. An important
step in preprocessing is stemming or lemmatizing. Stemming is a simple technique
where the ending of words is “chopped off”, leaving the stem. For example, using the
frequently used Porter stemming algorithm, “weaknesses” becomes “weak”, and both
“failures” and “failure” become “failur”. This technique does not handle irregular
conjugations, for example “are” and “were” have different stems. A more powerful (and
92 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
computationally demanding) technique is lemmatizing. Lemmatizing reduces all words
to their “lemma” using a lexicon in combination with regular conjugation rules. Thus,
lemmatization reduces both “is” and “were” to their lemma (to) “be”. For English,
stemming is often sufficient, but for more richly inflected languages such as German or
Dutch, lemmatization tends to give better results (Haselmayer and Jenny 2014).
Feature Selection
The next step is feature selection. A moderately large corpus can typically contain
more than 100,000 unique words, with most of these words occurring in only a few
documents. Applying an LDA model on all words in a corpus is both computationally
expensive and not very useful, as most words have distribution patterns that do not
contribute to meaningful topics. For example, some words are too frequent to be informative—words
like “the” and “have” generally occur in every document regardless of
topic. A useful technique to filter out words that are too rare or too common is to use
tf-idf (term frequency–inverse document frequency), which assigns a low score to words
that are either very rare or very frequent. Another option is to have a minimal
frequency cut-off to filter out the rare words and use a list of common stop words
(and/or cap the inverse document frequency) to filter out overly common words.
Lemmatization software is often combined with POS-tagging. POS (part of
speech) tags indicate the type of word, e.g. verb, noun or preposition. For topic modelling,
it is often best to only use specific parts of speech, especially nouns, proper
nouns and, depending on the task and corpus, adjectives and verbs. This automatically
filters out the most common stop words, which tend to be determiners or prepositions
(with the exception of common verbs like “to be” and “to have”).
For the analysis presented here, we used the lemmatizer and POS-tagger from
Stanford’s corenlp suite (De Marneffe, MacCartney, and Manning 2006), and selected all
nouns, verbs, adjectives and proper nouns. We filtered out all terms with a frequency
of less than 20 and which occurred in more than 25 per cent of documents, and we
removed all terms that contained a number or non-alphanumeric characters, yielding a
total vocabulary of 8493 terms.
Choosing Parameters
After creating a sufficiently small and relevant document–term matrix on which
to run the model, there are some more choices the researcher needs to make before
running the model. Although one of the main advantages of a topic model is that no a
priori coding schemes need to be supplied, there are certain parameters that need to
be set. In particular, the number of topics (K) needs to be specified, which indicates
into how many topics the LDA model should classify the words in the documents.
There is no default or simple rule of thumb for this parameter. The trade-off is comparable
to factor analysis: the goal is to describe the data with fewer dimensions (topics)
than are actually present, but with enough dimensions so that as little relevant information
as possible is lost. We first discuss and use the perplexity measure, which is a
commonly used computational indication for the correct amount of topics (Blei, Ng,
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 93
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
and Jordan 2003). However, we stress that this should be only used to make an initial
selection of models with an acceptable amount of information loss, and that
interpretability of topics is a more important criterion for social science purposes.
A second parameter is the alpha hyperparameter, which affects how many topics
a document can contain.2
A common default value for the alpha is 50 divided by the
number of topics. Substantively, a lower alpha leads to a higher concentration of topic
distributions within documents, meaning that documents score high on a few topics
rather than low on many. Accordingly, if the goal is to assign one or a few topics per
document then it makes sense to use a low alpha (Haselmayer and Marcelo 2014).
Perplexity
From a computational perspective, a good indication of the right number of
topics is that number with which the model best predicts the data. This is comparable
to goodness-of-fit measures for statistical models.3
For topic models such as LDA, a
commonly used indicator is perplexity, where a lower perplexity indicates a better prediction
(Blei, Ng, and Jordan 2003). To calculate the perplexity, we first train an LDA
model on a portion of the data. Then, the model is evaluated using the held-out data.
This routine is repeated for models with different numbers of topics, so that it becomes
clear which amount leads to the lowest perplexity.
Figure 2 shows the perplexity of different models for our data. We trained the
LDA models using 30,000 of the 48,604 documents, and then calculated the perplexity
of each model over the remaining 18,604 documents. We varied the number of topics
FIGURE 2
Perplexity of LDA models with different numbers of topics and alpha
Notes: The line graph shows how perplexity decreases (and model fit thus increases) as
the number of topics increases. The number of topics that corresponds to a great change
in the direction of the line graph is a good number to use for fitting a first model. For
example, K = 25 for our model with alpha = 50/K.
94 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
used: 2, 5, 10, 25, and then on to 250 in steps of 25. The results show that perplexity
decreases as the number of topics increases, implying that a model with more topics is
better at predicting the held-out data. At around 25 and 50 topics, the decrease in perplexity
for additional topics becomes notably less. This is one way to interpret the right
number of topics, similar to the interpretation of the elbow in the scree plot of a factor
analysis. The other way is to look at the number of topics with the lowest perplexity.
We can see that this point will be somewhere beyond 250.
However, having the “right” number of topics in a mathematical sense does not
say anything about the interpretability of the topics produced. In fact, mathematical
goodness-of-fit measures and human interpretation may lead to contradicting conclusions
on the best number of topics, given that especially with a high number of topics,
the computer algorithm may find nuances that are not semantically meaningful to
humans (Chang et al. 2009). Also, as we use topics to answer substantive questions
about the documents we study, it is important that the topics that result from the analysis
contribute towards answering these questions, instead of providing the best prediction
of the data. For the analysis presented in this article, we thus looked both at
perplexity and interpretability when deciding on the number of topics to use. Judging
from the perplexity, a good choice is probably between 25 and 50 topics. However, to
facilitate comparing our results with Gamson and Modigliani (1989), we will first use a
simpler model with K = 10 topics, and then show the differences between this model
and a model with K = 25 topics.
Alpha
Regarding the alpha hyper-parameter, since the data used for our analysis cover
a long history of textual data concerning an issue that involves various events and
viewpoints, it makes sense to define several clearly distinguishable topics. In addition
to better scores for the lower alpha, this is a substantive argument for us to use the
lower alpha of five divided by K instead of the default.
Tool Support for LDA
The easiest way to get started with LDA is through the open-source statistical
package R.4
Although specialized software for topic models is available, such as MALLET
(McCallum 2002) or the Stanford Topic Modeling Toolbox (Ramage et al. 2009), an
advantage of using R is that it is a statistical package that many social scientists already
use for other analyses. However, LDA works the same no matter the software used to
run the model. The tm package in R can automatically generate the document–term
matrix from texts and includes options for stemming and feature selection (Meyer, Hornik,
and Feinerer 2008). The topicmodels package can directly fit an LDA model from
the document–term matrix object created by tm. For more sophisticated preprocessing,
we use the Stanford corenlp suite (De Marneffe, MacCartney, and Manning 2006), which
contains a collection of modules for grammatical analysis. For our case study, we
uploaded our documents in the Web-based content analysis toolkit AmCAT.5
For the
preprocessing and the analysis itself, we used the statistical computing environment R,
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 95
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
that can connect to AmCAT directly to retrieve the documents there.6
AmCAT uses the
xtas extensible text annotation suite to automatically perform the preprocessing in a
scalable manner (De Rooij, Vishneuski, and De Rijke 2012).7
Nuclear Topics: 1945 to Now
In order to demonstrate the use of LDA to explore the topics in a given set of
documents, and show the change in these topics over time, we have performed LDA
on New York Times articles dealing with nuclear technology. The famous study by
Gamson and Modigliani (1989) shows how the framing of the issue “nuclear power” in
the news changed over time since 1945. We were interested to see whether the topics
found by performing an LDA over newspaper articles from this period show similar
changes over time. Additionally, we extended the research period to 2013, to include
more recent coverage of nuclear technology.
The question is thus whether the change in culture surrounding nuclear issues
found by Gamson and Modigliani is expressed in a change in word use over time that
is captured by LDA. Gamson and Modigliani (1989) identified seven “packages” or
frames in newspaper and television coverage of nuclear energy between 1945 and
1989: Progress, Energy Independence, Devil’s Bargain, Runaway, Public Accountability, Not
Cost Effective and Soft Paths (Gamson and Modigliani 1989, 24–25). If we compare the
outcome of an LDA analysis with the results of Gamson and Modigliani’s study, to what
extent do we find similar results? Note that for a number of reasons we do not expect
perfect correspondence between the LDA topics and Gamson and Modigliani’s packages,
even if we disregard the difference between manual and automatic analysis. First,
these packages were identified by examining not only the text of news stories, but also
images and cartoons, with a focus on editorial content, whereas our analysis is performed
on the lead paragraphs of news stories only. Second, whereas Gamson and
Modigliani only analyse nuclear power, our investigation deals with the whole coverage
of nuclear technology, including nuclear weapons. Finally, our investigation covers the
post-Cold War period as well, while Gamson and Modigliani of course only analyse the
discourse until the 1980s. Nonetheless, it is interesting to compare our findings to
theirs since it shows to what extent the results of an automatic topic modelling
approach compare to those of a well-known and very thorough manual analysis.
As discussed above, for reasons of comparability to Gamson and Modigliani’s
seven packages, we will first focus on an analysis using 10 topics. Table 1 shows the
topics that resulted from this analysis, including our interpretation and the 10 words
that represent a topic most strongly. To facilitate interpretation we also made a topic
browser,8
which is a tool to interactively explore the results of a topic model (Gardner
et al. 2010). Our topic browser features two extra pieces of information in addition to
top words. One is the top documents assigned to a topic in which the topic words are
highlighted. The other is a semantic network, or semantic map, which visualizes the cooccurrence
of top words, thus showing topic coherence and facilitating interpretation.
For our discussion, we categorized the topics into: (1) topics that show a pattern in
their use over time; (2) topics that are more or less continuously present; and (3) three
topics that turned out to be irrelevant for our case study.
96 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
Topics That Show Some Pattern in Their Use Over Time
Firstly, we found a number of topics that have a strong temporal dimension, that
is, they are strongly present in the news in some years or decades, but not in others. In
that respect, these topics are most similar to the shifting packages found by Gamson
and Modigliani. Figure 3 shows the change in occurrence over time for the four topics
discussed below, which are discussed in chronological order.
News stories in which topic 1, which we labelled Research, deal with research on
nuclear technology, including both energy research and nuclear weapons research. This
topic is most strongly present in the early part of the data-set, and its usage sharply
decreases over time, especially from the 1980s onwards. In terms of temporal focus,
and in the focus on possibilities created by nuclear research, this topic is comparable
to Gamson and Modigliani’s Progress package, although the latter did not include
nuclear weapons research as their research was focused on nuclear energy.
In topic 3, Cold War, the words “United”, “States”, “Soviet”, “Union” and “weapon”
immediately suggest that this is a topic about the US–Soviet conflict. Lower-ranking
words include variations on the “weapon” theme as well as more diplomatic terms such
as “agreement” and “proposal”. The topic occurs most frequently between the mid-
1950s and the mid-1980s, with a peak in the early 1960s that can be easily identified as
the Cuban missile crisis.
Finally, the last topic in this category is topic 8, Nuclear Power– Accidents/Danger.
Although the top words in this cluster, “plant”, “power” and “reactor”, are not very
informative, peripheral words like “accident”, “safety” and “radioactive” show how to
TABLE 1
LDA results on US nuclear discourse, 10 topics
Topic Interpretation Most representative words
Topics with temporal patterns
1 Research atomic, Energy, WASHINGTON, scientist, energy, bomb, Commission,
United, research, weapon
3 Cold War United, States, Union, Soviet, soviet, weapon, arm, missile, President,
treaty
7 Proliferation Iran, United, North, Korea, program, weapon, States, official, country,
China
8 Accidents/
Danger
plant, power, reactor, Island, Nuclear, accident, Commission, official,
waste, safety
Continuous topics
5 Weapons test, submarine, Japan, first, Navy, year, explosion, missile, ship, bomb
9 Nuclear
Power
power, plant, company, year, energy, percent, utility, cost, Company,
reactor
10 US Politics war, President, weapon, Mr., year, military, policy, world, Reagan,
House
Irrelevant topics
2 Summaries New, new, year, government, official, York, people, business, President,
state
4 Book
Reviews
week, life, book, man, woman, John, year, New, family, University
6 Films &
Music
Street, West, Theater, Mr., Sunday, East, show, New, tomorrow, p.m.
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 97
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
interpret it. Indeed, Chernobyl and Three Mile Island are both included in this topic,
which shows peaks in news attention in both 1979 and 1986. Articles on smaller
nuclear accidents or discussions on nuclear hazard in general also contain this topic.
Interestingly, although this topic shows a small peak around the 2011 Fukushima reactor,
most of the coverage of that event is classified in topic 5 (nuclear weapons), discussed
below. This topic is most closely related to Gamson and Modigliani’s Runaway
and Devil’s Bargain packages, as it focuses on the negative qualities of nuclear energy,
and both the topic and the Runaway frame occur in roughly the same time period.
Finally, the articles in topic 7, Nuclear Proliferation, deal with nuclear weapons in
countries such as North Korea, Iran and Pakistan, and US actions or policies against the
possession of these weapons. Peaks of attention for these topics occur in the 1990s
and especially the 2000s, with events in North Korea and later Iran as triggers. Although
the topic occurs after the period covered by Gamson and Modigliani, semantically it
resembles their Runaway package, in the sense that the technology is no longer under
control and now poses a danger to its very inventors.
More or Less Continuously Present Topics
These topics do have some fluctuation in their occurrence over time, but this
fluctuation shows no clear trend.
In topic 5, Nuclear Weapons, the words associated with the topic seem to be
mostly related to the development and testing of nuclear bombs, especially nuclear
submarines. This topic is most strongly present in the 1950s through the early 1970s,
with a peak in 2011 after the Fukushima incident. This latter peak is possibly a confusion
caused by the prevalence of ocean-related words in both the tsunami coverage
and the discourse on nuclear submarines.
FIGURE 3
Occurrence of topics that have a strong temporal component
98 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
Topic 9, Nuclear Power, is a second cluster with words related to economics.
Similar to topic 8, “power” and “plant” form the top words of a cluster. However, this
time the peripheral words show a different focus: “company”, “corporation”, “year” and
“percent” suggest that these articles are about nuclear energy as a business, and report
on the earnings of companies that operate in this sector. Over time, we see a peak in
the 1970s (the oil crisis) and in the 1980s.
The final relevant topic, topic 10, US Politics, concerns the policies of the United
States on its own nuclear weapons and defence. The main peak for this topic is in the
early 1980s, with the Rearming America programme.
Irrelevant Topics
These topics have nothing to do with nuclear power in terms of their content,
but appear in our results anyway because we included all articles mentioning the word
“nuclear”, including book reviews and news summaries. In topic 2, news summaries and
items from the section “Inside The Times” are clustered together as one topic. These
tend to focus on New York and on politics (local, domestic or foreign). However, they
do not deal with nuclear issues directly, so we discard this topic. Topic 4 consists of
short book reviews, of which one or more use words related to the nuclear topic.
Again, these do not deal with nuclear issues in the news directly. Lastly, topic 6 represents
news stories on films or concerts, some of which have to do with nuclear power,
others show up in articles that mention the nuclear issue elsewhere.
As this overview shows, not all topics contain useful results and the topics are
not ordered in a way that makes it easy to distinguish the useful from the non-useful
(as usefulness is of course something determined by the researcher, not the computer).
However, the irrelevant topics were clearly distinguishable, which makes it easy to discard
them altogether or ignore them in further analysis. Although it might seem annoying
at first that such topics are also generated by the analysis, this is actually quite
useful. Since most data-sets contain a degree of noise (such as book reviews or sports
results), LDA can be used as an inexpensive way to locate those articles that are relevant
for answering a particular research question from a larger sample.
Our purpose here, however, was to compare the topics we found using LDA with
the outcome of the framing research by Gamson and Modigliani (1989). Compared to
the frames or interpretive packages found in their analysis, the topics of our LDA analysis
seem to be more concrete and specific. Similar to Gamson and Modigliani’s study,
we found that topics change over time, but not all of them increase or decrease in a
linear way. Also, topics seem to either cluster a number of related events together (nuclear
proliferation talks, nuclear accidents) or represent issues that are continuously present
over a longer period of time (the economics of nuclear energy, nuclear weapons
tests). It is not possible, however, to deduct a particular viewpoint or frame from a
topic directly—for example, we found no clear “anti-nuclear” cluster, whereas Gamson
and Modigliani found multiple frames that are critical of nuclear energy. It is quite likely
that the coverage of nuclear accidents and danger is predominantly covered from an
anti-nuclear perspective, but even in this case there is a clear difference between the
“issue” or event being covered (nuclear accidents) and the frame with which it is
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 99
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
covered. That said, for some of the topics, such as the Research and Accidents topics,
we do see that the temporal pattern is similar to that identified for the Progress and
Runaway packages, respectively.
To see whether increasing the number of clusters helps find word patterns that
are more fine-grained and more frame-like than those representing issues, we will
explore what happens if we increase the number of topics from 10 to 25.
Granularity: How Many Topics?
Changing the number of topics (K) changes what is called the granularity, or level
of detail, of the model. The higher the granularity, the more detailed the analysis. Using
a larger number of topics implies higher granularity: each topic then represents more
specific content characteristics.
If a model with 25 topics is compared to a model with 10 topics, some of the
topics in the model with 25 topics tend to blend together in the 10-topic model. The
topics in the model with 50 topics can thus be considered to represent smaller grains
of the 10 topics. As an example, consider the events in Chernobyl and Three Mile
Island. Both events are clearly separated by time, space, the actors involved and various
circumstances. Yet the two are related by the nature of the event, and thus in vocabulary
used: a malfunction of a nuclear reactor, with awful consequences to health and
the environment. In a model with many topics, these events can be distinguished in
different topics, whereas in a model with fewer topics, these events are blended
together, representing a broader theme. In our data, we saw this specific example in a
comparison of models with 10 and 25 topics.
Table 2 gives an overview of the relevant topics from the 25-topic model. Each
“detailed” (K = 25) topic is listed below the “broader” (K = 10) topic it resembles most.
This similarity is computed by determining in which documents each K = 10 and
K = 25 topic occurs, and then calculating the cosine similarity of these occurrence
vectors. So, two topics with perfect similarity (1.0) would occur in exactly the same
(relative) frequency in all documents. For instance, topic 24 which we labelled 3 Mile
Island, is most similar to topic 8 above it (with cosine similarity 0.70).
In the K = 10 model, coverage of Chernobyl is clustered together with coverage
of the Three Mile Island incident, to form a cluster about nuclear accidents (topic 8),
and coverage of Fukushima was split between this topic and the nuclear weapons topic
(topic 5). However, in the K = 25 model, different connections between these events
are found, and Chernobyl and Fukushima end up in the same cluster, while Three Mile
Island gets its own cluster. Finally, topic 8 (K = 10) is highly similar to topic 15 (K = 25),
which deals with power plant construction and especially the Shoreham nuclear power
station. This power station was constructed on Long Island in the 1980s, but never
actually used, as local residents objected in the wake of the Chernobyl disaster. As
shown in Figure 4, these three constituent topics in the K = 25 model together trace
the K = 10 Accidents/Danger topic (the per-year correlation between the K = 10 topic
and the sum of the K = 25 topics is 0.95, p < 0.001), but each are focused around a
specific time-point.
100 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
Discussion: Best Practices
This article showcased a relatively recent tool, Latent Dirichlet Analysis, or LDA.
LDA is an unsupervised topic modelling technique that automatically creates “topics”,
that is, clusters of words, from a collection of documents. These topics may represent
issues that recur over time, related events or other regularities in articles.
We think LDA can be a valuable tool in any large-scale content analysis project.
For preliminary analysis, LDA can very quickly give a rough overview of what kind of
topics are discussed in which media or time periods. However, the best way of proving
the statistical, internal and external validity of LDA and of topic models in general is still
under discussion (Chang et al. 2009; DiMaggio, Nag, and Blei 2013; Ramirez et al. 2012).
We would advise journalism scholars to start by making a perplexity plot for different
numbers of topics, and then look at the models where the perplexity decrease drops
TABLE 2
Most similar K = 25 topics for each relevant K = 10 topic
Topic Similarity
1: Research
9: Research 0.68
8: Universities 0.39
17: Scientific Development 0.27
21: Nuclear Weapon Materials
3: Cold War
7: Rearming America 0.61
18: Cold War 0.61
20: NATO 0.50
8: Accidents/Danger
24: 3 Mile Island 0.70
15: Power Plant Construction (Shoreham) 0.62
2: Chernobyl & Fukushima 0.32
7: Nuclear Proliferation
6: Iran 0.63
4: North Korea 0.58
14: Iraq 0.42
3: India & Pakistan 0.25
21: Fissile Materials 0.25
5: Nuclear Weapons
11: Nuclear Submarines 0.53
2: Chernobyl & Fukushima 0.39
19: Nuclear Weapons 0.31
23: Nuclear Power 0.52
10: US Politics
1: Nuclear War Threat 0.68
5: US Politics 0.46
19: Nuclear Weapons 0.36
12: Protests 0.28
Topics from the 10-topic model (K = 10) are set in italics, with the most similar topics from the
25 topic model (K = 25) listed beneath each K = 10 topic. Some K = 25 topics occur twice
when they were similar to more than one K = 10 topic. Similarity is based on whether topics
occur in the same documents (calculated as cosine similarity). Example: Topic 3 from the
K = 10 model, interpreted as the Cold War topic, is similar to three topics from the K = 25
model: topic 7 on Reagan and Rearming America (sparking the 1980s’ arms race); topic 18 on
the Cold War; and topic 20 on NATO.
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 101
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
off. However, the researcher should also manually inspect these topic models, and for
each topic decide on the correct interpretation by looking at the words in the topic, in
which media and time frames it occurs, and also at the documents that are most
indicative of a specific topic. For this purpose, we advise using a topic browser.9
Manual
interpretation reveals which topics are closely related to the theoretical quantities
the researcher is interested in. Also, the researcher can decide to combine multiple
topics that are semantically related and/or to remove irrelevant topics such as the book
reviews identified in our case study. This should be followed by a formal internal validity
evaluation of these topics by checking a sample of automatically coded articles or
by comparing to a sample of manually coded articles (for a more elaborate discussion
of validity checks of LDA topic models, see DiMaggio, Nag, and Blei (2013). Even
though the validity may be insufficient for completely automatic analysis, it can be
used as a form of semi-automatic coding by having the manual coder check whether
the coding is correct, which is much quicker (and cheaper) than fully coding each document.
Furthermore, even if the topics cannot be immediately used to answer the substantive
research question, they can enhance the subsequent manual or automatic
content analysis in numerous ways. First, by inductively showing which topics occur, it
can help the researcher to create or improve the codebook by suggesting new codes
and examples and by showing which codes might need to be combined or split up.
Second, if the researcher is interested in doing automatic keyword analysis, the word
lists per topic can offer inspiration for keywords and synonyms that might otherwise be
left out. Third, LDA can be used to quickly find good example documents for infrequent
topics or documents that use multiple topics, such as the example document in
Figure 1. This can be used for manual coder training and evaluating the codebook, but
also for creating the training data for subsequent machine learning, where rare topics
usually give the worst performance because of the lack of training documents. Finally,
as shown by the three “irrelevant” topics that were derived from our analysis, LDA can
be used to filter out categories of texts that are not relevant from an overall sample.
FIGURE 4
Occurrence over time of detailed (K = 25) topics that constitute the Accidents/ Danger
topic from the K = 10 model
102 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
The comparison in this paper also shows two limitations of LDA for analysing
journalistic texts. First, not all topics represent substantive word clusters, but also other
consistent patterns of co-occurrence such as genre and writing style. This is most dramatically
shown by “irrelevant” topics such as book reviews, but can be considered
beneficial by allowing the researcher to quickly discard such clusters of documents.
However, it is also possible that shifts in word use, such as from “atomic” to “nuclear”,
caused documents to belong to different topics in the 1950s as compared to the 2000s
even though they are substantively similar. The same would hold for writing style differences
between different media and especially different formats (e.g. print media versus
television or online). Dynamic Topic Models (Blei and Lafferty 2006) and Structural
Topic Models (Roberts et al. 2014) are two extensions of LDA that explicitly deal with
shifts over time and between groups of documents, respectively.
Second, the topics identified in this study did not approach what we could call a
“frame” in the sense that Gamson and Modigliani (1989) or Entman (1993) would use
this concept—as coherent interpretative packages. Similarly, the topics did not represent
explicit valence or sentiment—there were no clear pro- or anti-nuclear topics.
Although attempts to combine LDA and sentiment analysis in one model have been
made (Lin and Yulan 2009), since news coverage is difficult for automatic sentiment
analysis in general (Balahur et al. 2013), such methods will likely not yield sufficiently
valid results for journalism studies in the near future. Following DiMaggio, Nag, and Blei
(2013, 603), we suggest further research can combine LDA as described here with frame
and sentiment analysis using other methods, e.g. using machine learning (e.g. Burscher
et al. 2014).
By showcasing LDA and by showing some best practices for running and interpreting
model results, this article contributes to the adoption of topic modelling in the
practice of journalism research, which is a useful technique for every digital journalism
scholar to have in their toolbox to deal with the very large data-sets that are becoming
available. Although the burden of making sense of the results is still on the researcher,
LDA offers a quick and cheap way to get a good overview of a collection of documents
so that substantive questions can be answered immediately, especially about broad
patterns in topic use over time or between media. Additionally, it is very helpful for
performing preliminary analysis before venturing on a more traditional (and expensive)
automatic or manual content analysis project.
DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.
NOTES
1. Factor analysis is a dimensionality reduction technique: given a set of observed
variables, a smaller set of factors is calculated that preserve as much information
as possible in a lower-dimensional space. This is often used in the field of psychology
as a measurement of latent, unobserved causes for certain observations.
For instance, if a single factor largely explains the results for a set of questions
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 103
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
relating to anxiety, the factor can be interpreted as a measurement of anxiety.
Similarly, a topic in topic modelling can be interpreted and named based on what
the main words have in common.
2. Technically, the alpha hyper-parameter controls the concentration of the Dirichlet
distribution regarding the distribution of topics over documents. In Bayesian
statistics, a hyper-parameter is a parameter that controls distributions such as the
Dirichlet distribution. The term hyper-parameter is used to distinguish them from
the parameters of the topic model that is the result of the analysis. For a good
explanation of the role of hyper-parameters, we suggest the introduction to the
Dirichlet distribution by Frigyik, Kapila, and Gupta (2010).
3. A goodness-of-fit measure describes how similar the predicted or expected values
of a model are to the actual observed values. An example is the R2
measure in
linear regression, which indicates what proportion of variance of the dependent
variable is explained by the independent variables.
4. See http://www.r-project.org.
5. See http://amcat.nl.
6. See http://github.com/amcat/amcat-r for the relevant R code. The R scripts that
were used for our analysis can be downloaded from http://github.com/
AUTHOR/corpus-tools.
7. See http://github.com/AUTHOR/xtas for the xtas modules for corenlp and other
lemmatizers.
8. The topic browser can be found at http://rpubs.com/Anonymous/78,706.
9. Our R script for creating a topic browser is available at http://github.com/vanat
teveldt/topicbrowser.
REFERENCES
Balahur, Alexandra, Ralf Steinberger, Mijail Kabadjov, Vanni Zavarella, Erik Van Der Goot,
Matina Halkia, Bruno Pouliquen, and Jenya Belyaeva. 2013. “Sentiment Analysis in the
News.”Proceedings of the 7th International Conference on Language Resources and
Evaluation (LREC’2010): 2216-2220. Valletta, 19–21 May 2010.
Blei, David. M., and John D. Lafferty. 2006. “Dynamic Topic Models.” In Proceedings of the
23rd International Conference on Machine Learning, 113–120. Pittsburgh, 2006.
doi: 10.1145/1143844.1143859.
Blei, David M., Andrew Y. Ng, and Michael I. Jordan. 2003. “Latent Dirichlet Allocation.” The
Journal of Machine Learning Research 3: 993–1022.
Burscher, Bjo¨rn, Daan Odijk, Rens Vliegenthart, Maarten de Rijke, and Claes H. de Vreese.
2014. “Teaching the Computer to Code Frames in News: Comparing Two Supervised
Machine Learning Approaches to Frame Analysis.” Communication Methods and
Measures 8 (3): 190–206. doi: 10.1080/19312458.2014.937527.
Chang, Jonathan, Sean Gerrish, Chong Wang, Jordan Boyd-Graber, and David M. Blei. 2009.
“Reading Tea Leaves: How Humans Interpret Topic Models.” In Advances in Neural
Information Processing Systems, 288–296.
De Marneffe, Marie-Catherine, Bill MacCartney, and Christopher D. Manning. 2006. “Generating
Typed Dependency Parses from Phrase Structure Parses.” In Proceedings of Lrec 6:
449–454.
104 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
De Rooij, Ork, Andrei Vishneuski, and Maarten De Rijke. 2012. “Xtas: Text Analysis in a Timely
Manner.” Paper presented at Dir 2012: 12th Dutch-Belgian Information Retrieval
Workshop. Ghent, 23–24 February 2012.
DiMaggio, Paul, Manish Nag, and David Blei. 2013. “Exploiting Affinities between Topic
Modelling and the Sociological Perspective on Culture: Application to Newspaper
Coverage of U.S. Government Arts Funding.” Poetics 41:570–606.
Entman, Robert M. 1993. “Framing: Toward Clarification of a Fractured Paradigm.” Journal of
Communication 43 (4): 51–58.
Frigyik, Bela A., Amol Kapila, and Maya R. Gupta. 2010. “Introduction to the Dirichlet Distribution
and Related Processes”. Technical Report UWEETR-2010-0006. Seattle: University
of Washington Department of Electrical Engineering.
Gamson, William A., and Andre Modigliani. 1989. “Media Discourse and Public Opinion on
Nuclear Power: A Constructionist Approach.” American Journal of Sociology 95: 1–37.
Gardner, Matthew J., Joshua Lutes, Jeff Lund, Josh Hansen, Dan Walker, Eric Ringger, and
Kevin Seppi. 2010. “The Topic Browser: An Interactive Tool for Browsing Topic
Models.” In Proceedings of the Neural Information Processing Systems (NIPS) Workshop
on Challenges of Data Visualization: 5528–5235. Whistler, December 11, 2010.
Grimmer, Justin, and Brandon M. Stewart. 2013. “Text as Data: The Promise and Pitfalls of
Automatic Content Analysis Methods for Political Texts.” Political Analysis 21 (3):
267–297. doi:10.1093/pan/mps028.
Haselmayer, Martin, and Marcelo Jenny. 2014. “Measuring the Tonality of Negative
Campaigning: Combining a Dictionary Approach with Crowd-Coding.” Paper presented
at Political Context Matters: Content Analysis in the Social Sciences. University of
Mannheim.
Jones, Bryan D., and Frank R. Baumgartner. 2005. The Politics of Attention: How Government
Prioritizes Problems. Chicago, IL: University of Chicago Press.
Lin, Chenghua, and Yulan He. 2009. “Joint Sentiment/Topic Model for Sentiment Analysis.” In
Proceedings of the 18th ACM conference on information and knowledge management
(5), 375–384. Hong Kong, November 2–6, 2009.
Lucas, Christopher, Richard A. Nielsen, Margaret E. Roberts, Brandon M. Stewart, Alex Storer,
and Dustin Tingley. 2015. “Computer-Assisted Text Analysis for Comparative Politics.”
Political Analysis 23: 254–277.
Matthes, Jo¨rg, and Matthias Kohring. 2008. “The Content Analysis of Media Frames: Toward
Improving Reliability and Validity.” Journal of Communication 58 (2): 258–279.
doi:10.1111/j.1460-2466.2008.00384.x.
McCallum, Andrew K. 2002. Mallet: A Machine Learning for Language Toolkit. Retrieved from
http://mallet.cs.umass.edu
Meyer, David, Kurt Hornik, and Ingo Feinerer. 2008. “Text Mining Infrastructure in R.” Journal
of Statistical Software 25 (5): 1–54.
Quinn, Kevin M., Burt L. Monroe, Michael Colaresi, Michael H. Crespin, and Dragomir R.
Radev. 2010. “How to Analyze Political Attention with Minimal Assumptions and
Costs.” American Journal of Political Science 54 (1): 209–228. doi:10.1111/j.1540-
5907.2009.00427.x.
Ramage, Daniel, Evan Rosen, Jason Chuang, Christopher D. Manning, and Daniel A.
McFarland. 2009. “Topic Modeling for the Social Sciences.” In Nips Workshop on
Applications for Topic Models: Text and beyond 5: 1–4.
ANALYSIS OF JOURNALISTIC TEXTS USING TOPIC MODELLING 105
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017
Ramirez, Eduardo H., Ramon Brena, Davide Magatti, and Fabio Stella. 2012. “Topic Model
Validation.” Neurocomputing 76 (1): 125–133.
Roberts, Margaret E., Brandon M. Stewart, Dustin Tingley, Christopher Lucas, Jetson
Leder-Luis, Shana K. Gadarian, B. Albertson, and David G. Rand. 2014. “Structural Topic
Models for Open-Ended Survey Responses.” American Journal of Political Science 58(4):
1064–1082. doi: 10.1111/ajps.12103.
Carina Jacobi (author to whom correspondence should be addressed), Department
of Communication, University of Vienna, Austria; E-mail: carina.jacobi@univie.
ac.at
Wouter van Atteveldt, Department of Communication Science, VU University Amsterdam,
The Netherlands; E-mail: wouter@vanatteveldt.com. http://vanat
teveldt.com
Kasper Welbers, Department of Communication Science, VU University Amsterdam,
The Netherlands; E-mail: k.welbers@vu.nl
106 CARINA JACOBI ET AL.
Downloadedby[MasarykovaUniverzitavBrne]at11:0918September2017