Big data analysis of news and social media content

Big Data Analysis of News and Social Media Content
Ilias Flaounas, Saatviga Sudhahar, Thomas Lansdall-Welfare,
Elena Hensiger, Nello Cristianini (*)
Intelligent Systems Laboratory, University of Bristol
(*) corresponding author
Abstract
The analysis of media content has been central in social sciences, due to the key role that
media plays in shaping public opinion. This kind of analysis typically relies on the
preliminary coding of the text being examined, a step that involves reading and annotating
it, and that limits the sizes of the corpora that can be analysed. The use of modern
technologies from Artificial Intelligence allows researchers to automate the process of
applying different codes in the same text. Computational technologies also enable the
automation of data collection, preparation, management and visualisation. This provides
opportunities for performing massive scale investigations, real time monitoring, and
system-level modelling of the global media system. The present article reviews the work
performed by the Intelligent Systems Laboratory in Bristol University towards this direction.
We describe how the analysis of Twitter content can reveal mood changes in entire
populations, how the political relations among US leaders can be extracted from large
corpora, how we can determine what news people really want to read, how gender-bias
and writing-style in articles change among different outlets, and what EU news outlets can
tell us about cultural similarities in Europe. Most importantly, this survey aims to
demonstrate some of the steps that can be automated, allowing researchers to access
macroscopic patterns that would be otherwise out of reach.
Introduction
The ready availability of masses of data and the means to exploit them is changing the
way we do science in many domains (Cristianini, 2010; Halevy et al., 2009). Molecular
biology, astronomy and chemistry have already been transformed by the data revolution,
undergoing what some consider an actual paradigm shift. In other domains, like social
sciences (Lazer et al., 2009; Michel et al., 2011) and humanities (Moretti, 2011), data-
driven approaches are just beginning to be deployed. This delay was caused by the
complexity of the social interactions and the unavailability of digital data (Watts, 2007).
Examples of works that deploy computational methods in social sciences include: the
study of human interactions using data from mobile phones (Onnela et al., 2007; González
et al., 2008); the study of human interactions in online games and environments (Szell et
al., 2010); or automating text annotation for political science research (Cardie et al., 2008).
In particular, the analysis of media content in the social sciences, has long been an
important concern, due to the important role that media play in reflecting and shaping
public opinion (Lewis, 2001). This includes the analysis of both traditional news outlets,
such as newspapers and broadcast media, and modern social media such as Twitter
where content is generated by users. The analysis of news content is traditionally based
on the coding of data by human coders. This is a labour-intensive process that limits the
scope and potential of this kind of investigation. Computational methods from text mining
and pattern recognition, originally developed for applications like e-commerce or

information retrieval, can be deployed to automate many aspects of data collection,
annotation, analysis and visualisation.
Our group at the Intelligent Systems Laboratory in Bristol has worked for the past six years
in this direction, creating technologies to support Social Sciences investigations and to
demonstrate their use on real-world large scale datasets. We have built a computational
infrastructure which is capable of gathering, storing, translating, annotating and visualising
vast amounts of news items and social media content (Flaounas et al., 2011). Currently
our system tracks 1,100 news outlets in 22 languages, and tweets from 54 UK locations.
Overall the system has generated over one Terabyte (equivalent to 1000 Gigabytes) of
data over the past four years.
This article reviews part of our research in media content analysis, which we call the
MediaPatterns project1
. The technical details have been avoided in an attempt to
emphasise the capabilities of data-driven approaches in social sciences research. For
each study we provide references to the original articles and websites where the results
were first presented. It is important to mention that each study was performed by using
different methodologies depending on the task. In the present article we will review the
following: Sentiment analysis of Twitter content (using keyword matching and time-series
analysis techniques); Narrative analysis of US Elections (using natural language
processing and network analysis techniques); Comparison of news outlets based on topics
covered, writing style, and gender bias (using machine learning and natural language
processing techniques); Modelling of reader preferences (using machine learning
techniques); Network analysis of the EU mediasphere (using clustering and network
analysis techniques); and event detection in textual streams (using statistical approaches).
Sentiment Analysis of Twitter Content
Measuring the current public mood is a challenging task. The traditional approach would
require questioning a large number of people about their feelings. Social media, such as
Twitter or Facebook, can easily become a valuable source of information about the public
due to the fact that people use them to express their feelings in public.
As demonstrated in our study by Lansdall-Welfare et al. (2012) it is feasible to capture the
public mood by monitoring the stream of Twitter data. The dataset that was analysed was
comprised of 484 million tweets that were generated by more than 9.8 million users,
between July 2009 and January 2012. The data were collected from the 54 largest cities in
the UK. We focused on tracking four moods which are “Fear”, “Joy”, “Anger” and
“Sadness”. For each mood, we track a long list of associated words and we count the
frequencies that these words appear in tweets. This process generates one timeline of the
volume of related tweets for each emotion. The further analysis of these timelines reveals
that each of the four emotions changes over time in a rather predictable manner. To give
an example, we found a periodic peak of joy around Christmas and a periodic peak of fear
around Halloween. More surprisingly, we found that negative moods started to dominate
the Twitter content after the announcement of massive cuts in public spending on October
2010. Also there was a significant increase in anger in the weeks before the summer riots
of August 2011. In Fig. 1 we plot the mood levels for the period of study and we visualise
them as a facial expression using the Grimace tool. In the project website we have an
1 MediaPatterns website: http://mediapatterns.enm.bris.ac.uk

interactive demo that can be used to explore our data2
.
A detailed interpretation of the results of this study is a challenging task. A scholar needs to
understand the causes of the measured statistical quantities and associate them with
specific events that affected society. The interpretation of the results can not be automated
by using computational methods and it is a task that needs to be undertaken by social
scientists.
Figure 1. Analysis of 484 million tweets over a period of three years can reveal changes in
public mood. We track four basic emotions and we can associate mood changes with
important events (Lansdall-Welfare et al., 2012).
Narrative Analysis of US Elections
Content analysis sometimes involves the identification of basic narrative information in a
corpus. For example, determining the key actors, the key actions, and how actors interact.
Quantitative Narrative Analysis (Franzosi, 1987; Earl et al., 2004) involves the process of
extracting Subject-Verb-Object (SVO) triplets and their study in order to analyse a corpus,
while “distant reading” in humanities transforms novels into networks as a first step of
analysis (Moretti, 2011). The extraction of SVO triplets and turning them into a network is a
process that can be automated and scaled up, enabling the analysis of big corpora.
In our work by Sudhahar et al. (2012), we used computational methods to perform
narrative analysis of the US Elections 2012. The dataset we analysed is comprised of
125,254 articles that were automatically categorised as being related to the US Elections
using the machine learning Support Vector Machines (SVM) approach (Cristianini, 2000).
We pre-processed the corpus text using GATE (Cunningham, 2002) tools and then we
used Minipar parser (Lin, 1998) to extract Subject-Verb-Object triplets. We identified
31,476 actors from these triplets. A statistical analysis identified the key actors and
actions. Furthermore, we separate actions into two categories, endorsement and
opposition. Finally, the key actors and actions were used to infer the network of the US
Elections. This network is presented in Fig. 2 were the protagonists of the Elections,
“Obama” and “Romney”, are easily observed to dominate the network.
2 Mood Changes in Twitter Content: http://mediapatterns.enm.bris.ac.uk/mood/

Exploring this network can give a useful insight on the topics that are most discussed
during the elections, who the most dominant or least reported people are, how often the
key actors are presented by media as endorsing or opposing topics etc. Our results can be
accessed online through an interactive website that we built for the monitoring of the
elections3
. The website is updated daily for the duration of the elections. A further analysis
would require a social scientist to find and deeply understand the causes and associate
them with raw data in order to interpret the patterns that emerge.
Figure 2. The key actors and their relations automatically extracted from a corpus of
125,254 articles related to US Elections 2012 (Saatviga et al. 2012).
Comparison of News Outlets on Topics, Writing Style and Gender Bias
Detection of differences or biases among news outlets is a popular field of research in
media studies. Nowadays, several aspects of content analysis can be automated: using
machine learning techniques such as Support Vector Machines (Cristianini et al., 2000) we
can analyse text and detect general topics, such as Politics or Sports, that are present in a
corpus; we can use natural language processing techniques to quantify properties of the
writing style of the text or detect mentions of people and their gender.
In our work by Flaounas et al. (2012), we analysed a large dataset comprised of 2.5 million
news articles collected from 498 different news outlets over a period of 10 months. For
each article we identified the general topics that it covered, the people that they were
mentioned and their gender (Cunningham et al., 2002), as well as two basic writing style
properties, namely the readability and linguistic subjectivity. Readability is a measure of
how easy it is to read a text calculated by taking into account the length of sentences and
the length of the words in syllables (Flesch, 1948). Linguistic subjectivity is a measure of
the usage of words that carry sentiment, i.e. the more words with sentiment present in an
article, the more linguistically subjective it is (Pang, 2008).
The computation of the aforementioned quantities allows different comparisons of sets of
articles. For example, for the articles of each topic we calculated: a) the average
readability - finding that articles about sports are the easiest to read while articles on
3 ElectionWatch: http://electionwatch.enm.bris.ac.uk

politics are the hardest to read; b) Linguistic subjectivity - finding that articles about fashion
and arts are the most subjective while business articles were the most objective; c) the
ratio of males over females – finding that articles about fashion and art mention the most
women while articles about sports are male dominated. Furthermore, we directly compared
15 major US and UK newspapers on which topics they tend cover more often; their writing
style and also the ratio of males over females that their articles mention. In Fig. 3 we
visualise the comparison of outlets based on their writing style: outlets with similar writing
style are closer together.
The approach for automatically detecting differences between outlets, people, topics,
countries etc. is straightforward. The combination of large amounts of data in digital format
with the application of machine learning and natural languages techniques can help social
scientists answer a diverse range of research questions that could not be posed before.
Figure 3. Newspapers compared based on their Readability and Linguistic Subjectivity.
The illustration is based on the analysis of a dataset comprised of articles from US and
UK newspapers (Flaounas et al. 2012).
The EU News Media Network
A recent technology that enables new avenues of research is machine translation. The
technology is not mature enough to create human quality translations but nevertheless
allows a scholar to access textual material written in most languages. Furthermore,
machine translation can be combined with other computational methods in order to
analyse texts written in non-English languages using computational tools developed for
English. This reduces the costs of developing specialised tools for each language
separately.
In our paper by Flaounas et al. (2010), we showed how machine translation allowed the
content analysis of a large corpus comprised of 1.3 million articles published by main news
media in the 27 EU countries over a period of six months. Those articles were written in 22
different European languages. All non-English articles were machine-translated into
English before further processing. In this research, we used the machine learning
approach of clustering to identify sets of articles that refer to the same event. Then

statistical approaches were used to infer the network of the EU news media. In that
network, two outlets are connected if they tend to publish articles about the same events.
In Fig. 4 we illustrate a sparse version of that network. It can be observed that news
outlets, represented as nodes of the network, are organised in communities. An interesting
and expected result is that outlets from the same country belong to the same community,
i.e. they are interested in the same stories.
A deeper analysis of the structure of the network revealed some less expected results. For
example, that the EU media system reflects the geographic, economic, and cultural
relations between the EU countries. Also, we showed that the deviation of the content of
an EU news outlet from the “average” EU media content is significantly correlated with the
year of accession to the EU of the country that the outlet is based (correlation 48.94%, p-
value<0.01). Interestingly enough, the editors of 215 different news outlets from 27
countries made their independent choices over what stories they will cover; but these
choices shape the EU media sphere in a way that reflects relations between EU countries.
The detection of these subtle signals in a statistically rigorous way required the analysis of
a massive dataset and the deployment of methodologies such as machine translation,
data clustering and network analysis. To conclude, we showed that ideally, the language
differences should not be an obstacle for a scholar to formulate research questions and
analyse data.
Figure 4. The communities of the network of EU news outlets. Nodes represent outlets
and two outlets are connected if they tend to cover the same stories. Outlets from the
same country are coloured with the same colour. To infer this network we used more
than 1.3 million articles written in 22 different languages (Flaounas et al., 2010).
What Makes Us Click
Some online news outlets highlight their “Most Popular” stories. This is a valuable source
of information since it can be used to model the reading preferences of their audience. The
modelling starts by forming a large number of pairs of articles. Each pair is comprised of
one popular article and one non-popular article, i.e. an article that was published on the
main page of the outlet, on the same day, but it didn't become popular. Then a machine
learning algorithm, namely Ranking-SVM, can be used to create a statistical model that
captures the words that tend to appear in popular articles (Hensinger et al., 2012a). The
same model can be used to predict if any article will become popular or not.

In the study by Hensinger et al. (2012b), we showed how we can model the reading
preferences of the audiences of 14 news outlets. We collected popular and non-popular
articles over a period of 20 months forming in total 2,432,148 pairs of articles. For each of
the 14 news outlets we created a model of the preferences of their readers. These models
had an average performance of 70.6% in predicting which article between a pair of articles
has a better chance of becoming popular. In Fig. 5 we present as an example the word
cloud of the model for “News.com.au” which shows the words that trigger the attention of
the audience and the words that contribute less to the appeal of an article.
Similar approaches to the one we present can also be used to identify the preferences of
people, by monitoring their online behaviour - simply checking what they click on. Both
approaches offer a coarse understanding of people's behaviour. Of course, a deeper
understanding would require a more traditional approach of questioning and interviewing
people and reasoning about their answers.
Figure 5. The word cloud presents in pink the words that are present in high-appeal
articles published by “News.com.au” and in black the words that are present in low-appeal
articles. The size of each word reflects the contribution of the word, either positive or
negative, to appeal. The analysed dataset contained 2.4 million of pairs of popular and
non-popular articles (Hensinger et al. 2012b).
Event Detection in Textual Streams
There are various ways in which big-data can be found “in the wild” and one increasingly
important format is that of textual data streams, such as those generated by news feeds,
social media or personal communications. When monitoring a data stream, an important
issue is detecting change points. These are any variations in the statistical properties of its
content that might signal important events.
In the research by Snowsill et al. (2010), we demonstrated one possible approach based
on tracking sequences of words (n-grams) using a suffix tree methodology. We applied it
to the New York Times corpus on articles published during a period of 10 years, from
January 1996 until June 2007. The dataset we analysed contained in total 1,035,263
articles. Using a statistical approach we detected and ranked the major events. We plotted

a timeline for each event for the whole period of the 10 years we analysed. In Fig. 6 we
illustrate a selection of events. For example, detected events include “Princess Diana”,
“Iraq”, “World Trade Center”, and “Afghanistan”. Note, that these major events were
detected in a fully automated way and no-human supervision or interaction was necessary
at any point during the whole process.
Once more the interpretation of these results, their deeper understanding and the
extraction of useful knowledge out of them is a task for a social scientist. The
computational approach can only highlight where the scholar should focus their attention.
Figure 6. We analysed 1,035,263 articles from the New York Times corpus published over
a period of ten years and we detected the most significant events reported (Snowsill et al.,
2010).
Conclusions
The capability to gather vast amounts of digital data and the computational methodologies
to analyse them are changing many aspects of science. The research questions that can
be answered using data-driven approaches evolve and are multiplied based on the
progress of pattern-analysis and data-mining technology. Machines provide the
opportunities for extracting valuable information from big data and performing the repetitive
and mundane tasks. Their capabilities to digest large amounts of data, far more than any
researcher could analyse, are compensating for their inaccuracies and lack of reasoning.
Of course, in many cases there is no substitute for human interpretation and judgement.
The boundary between tasks that can be solved by machines and those that can not is
constantly changing. This is particularly true in the case of text analysis technologies. For
example, just a decade ago machine translation could not be listed among the tools
available to social scientists. The technologies of text-mining and text-analysis are evolving
quickly, driven by strong commercial pressure, and as a side effect can lead to new tools
for the social sciences. It is certainly worth keeping an eye on which new tools become
available, as they could provide help in answering entirely new research questions. At the
same time, entirely new types of data are constantly being released, ranging from news to
social media content, from political proceedings to the content of millions of books (Michel
et al., 2011). The opportunities for progress are real, when we properly combine these two
types of resources.

The work surveyed in this paper demonstrates technologies that can reliably detect topics,
sentiment, persons, events and relations among actors, applied on millions of documents.
Computational methods can also retrieve them, translate them and code them, all on a
scale that cannot be matched by human analysts. We followed common practices in text
analysis, by calibrating the tools we deployed on reference datasets that have been
annotated by humans, so that we know the precision of our automatic annotation. For
example to calibrate topic detection algorithms we used the public datasets of New York
Times (Sandhaus, 2008) and the Reuters (Lewis et al, 2004), while for discovering the
gender of people we used the public database Freebase4
. While this is clearly not perfect,
there are applications for which this is sufficient. It is important to notice that data
annotated by human researchers during their own studies can provide a very valuable
resource for the training and evaluation of artificial tools, and should be shared and be
open, much like is done in biology (Let data speak to data, 2005).
One important aspect of using big data in research has been mentioned only marginally in
this survey: the need to visualise the data and the relations discovered in it, in order to
help analysts make sense of the large scale datasets. We have demonstrated how political
relations can be represented as networks, how mood can be represented as a face
expression, but much more can be done in this direction.
While we have focused on the analysis of textual news content, and any other structure we
derived from it (e.g. networks, maps and timelines), many researchers in the area of
Computational Social Science (CSS) focus on different types of data: for example social
networks, or search engine queries. In most of those cases, all data from CSS share the
feature of being mostly unstructured, that is not naturally organised in rigorous structures,
which is instead the norm in engineering. The exploitation of unstructured data is
becoming a major area of concern in computer science and we can expect fast progress in
the next few years.
As we move quickly towards a new way of doing science, we can see both opportunities
and risks. It is important to keep both of them in mind, and to avoid unnecessary hype.
There are many cases where forcing the culture of engineering on a domain such as the
social sciences, but also the natural sciences, would be a mistake. We should keep in
mind that these tools can only hope to complement – never to replace – the role of
analysts, with their capability to understand context, to interpret and to incorporate their
findings into a coherent narration. Data-driven technologies can be useful when they allow
analysts to become aware of macroscopic trends that would be otherwise out of reach, but
we should not delude ourselves that they have any chance of ever taking their place.
Acknowledgments
I. Flaounas and N. Cristianini are supported by European Community's Seventh
Framework Programme under grant agreement N° 270327 (CompLACS Project). Authors
would like to acknowledge the former members of the group, namely Omar Ali, Vasileios
Lampos, Tristan Snowsill, and Marco Turchi for their contributions to the publications that
were reviewed in this paper.
References
Let data speak to data (2005), (Editorial), Nature 438, pp. 531.
4 Freebase: http://www.freebase.com

Cardie, C, Wilkerson, J (2008) “Text annotation for political science research”, Journal of
Information Technology & Politics, 5(1), pp. 1–6.
Cristianini, N, Shawe-Taylor, J (2000) “An Introduction to Support Vector Machines and
other Kernel-based learning methods”, Cambridge University Press.
Cristianini, N (2010), “Are we there yet?”, Neural Networks, Elsevier, 23(4), pp. 466-470.
Cunningham H, Maynard D, Bontcheva, K, And Tablan, V (2002) “GATE: A Framework and
Graphical Development Environment for Robust NLP Tools and Applications”, In Proc. of
the 40th Anniversary Meeting of the Association for Computational Linguistics, pp. 168–
175.
Earl J, Martin A, McCarthy J, Soule S (2004) The use of newspaper data in the study of
collective action. Annual Review of Sociology 30:65–80.
Halevy, A, Norvig, P and Pereira, F (2009), “The unreasonable effectiveness of data”,
Intelligent Systems, IEEE, vol. 24, pp. 8-12.
Flaounas, I, Turchi, M, Ali, O, Fyson, N, Bie, TD, Mosdell, N, Lewis, J & Cristianini, N
(2010), 'The Structure of EU Mediasphere' PLoS ONE, vol 5 (12).
Flaounas, I, Ali, O, Turchi, M, Snowsill, T, Nicart, F, De Bie, T, Cristianini, N (2011) “NOAM:
news outlets analysis and monitoring system”, SIGMOD Conference, ACM, pp. 1275-
1278.
Flaounas, I, Ali, O, Lansdall-Welfare, T, De Bie, T, Mosdell, N, Lewis, J & Cristianini, N
(2012), 'Research Methods in the Age of Digital Journalism: Massive-scale automated
analysis of news-content: topics, style and gender', Digital Journalism, Routledge, vol 1,
no. 1.
Flesch, R (1948) “A New Readability Yardstick”, Journal of Applied Psychology 32, pp.
221–233.
Franzosi R (2010) Quantitative Narrative Analysis. Sage Publications Inc, Quantitative
Applications in the Social Sciences, pp. 162:200.
González, M, Hidalgo, C, and Barabási, A (2008), Understanding individual human mobility
patterns. Nature, 453(7196), pp. 779–782.
Hensinger, E, Flaounas, I, Cristianini, N (2012a) “What makes us click? Modelling and
Predicting the Appeal of News Articles”, Proc. of International Conference on Pattern
Recognition Applications and Methods, pp. 41-50.
Hensinger, E, Flaounas, I, Cristianini, N (2012b) “The Appeal of Politics on Online
Readers”, In: Internet, Politics, Policy 2012:Big Data, Big Challenges?, Oxford, Available
at: http://microsites.oii.ox.ac.uk/ipp2012 [Accessed 1st
Nov. 2012].
Lansdall-Welfare, T, Lampos, V & Cristianini, N (2012), 'Effects of the Recession on Public
Mood in the UK'. In Mining Social Network Dynamics (MSND) session on Social Media
Applications in News and Entertainment at WWW '12., ACM, pp. 1221 – 1226.
Lazer, D, Pentland, A, Adamic, L, Aral, S, Barabasi, A-L, Brewer, D, Christakis, N, et al.
(2009) “Computational Social Science”, science 323, pp. 721-723.
Lewis, J (2001) “Constructing Public Opinion, New York: Columbia University Press.
Lewis, D, Yang, Y, Rose, T G, and Li, F (2004) “RCV1: A New Benchmark Collection for
Text Categorization Research”, Journal of Machine Learning Research 5, pp. 361–397.
Lin D, (1998), “Dependency-Based Evaluation of Minipar”, Text, Speech and Language
Technology 20 pp. 317–329.
Michel, J-B, Shen, Y K, Aiden, A P, Veres, A, Gray, M K., et al. (2011) “Quantitative

Analysis of Culture Using Millions of Digitized Books”, Science 331, pp. 176–182.
Moretti, F (2011) “Network theory, plot analysis”. New Left Review 68, pp. 80–102.
Onnela, J, Saramäki, J, Hyvönen, J, Szabó, G, Lazer, D, Kaski, K, Kertész, J, and
Barabási, A (2007) “Structure and tie strengths in mobile communication networks”,
Proceedings of the National Academy of Sciences, 104(18), pp. 7332–7336.
Pang, B, Lee, L (2008) “Opinion mining and sentiment analysis”, Foundations and Trends
in Information Retrieval 2, pp. 1–135.
Sandhaus, E (2008) “The New York Times Annotated Corpus”, The New York Times
Company, Research and Development.
Snowsill, T, Flaounas, I, De Bie, T, Cristianini, N (2010), 'Detecting events in a million New
York Times articles'. In Proc. of Machine Learning and Knowledge Discovery in Databases,
Springer, pp. 615-618.
Sudhahar, S, Lansdall-Welfare, T, Flaounas, I, Cristianini, N (2012), 'ElectionWatch:
Detecting Patterns In News Coverage of US Elections'. In Proceedings of the 13th
Conference of the European Chapter of the Association for Computational Linguistics
(EACL), pp. 82-86.
Szell, M, Lambiotte, R, Thurner, S (2010) “Multirelational organization of large-scale social
networks in an online world”, Proceedings of the National Academy of Sciences, 107(31),
pp. 13636–13641.
Watts, D (2011), “Everything is Obvious: Once You Know the Answer: How Common
Sense Fails”, Crown Business, New York.
Watts, D (2007) “A twenty-first century science”, Nature 445, pp. 489.

Big data analysis of news and social media content

More Related Content

What's hot (20)

Similar to Big data analysis of news and social media content (20)

More from Firas Husseini (20)

Recently uploaded (20)

Big data analysis of news and social media content