In this paper we address the problem of identifying attention dominating moments in online media. We are interested in discovering moments when everyone seems to be talking about the same thing. We investigate one particular aspect of breaking news: the tendency of multiple sources to concentrate attention on a single topic, leading to a collapse in diversity of content for a period of time. In this work we show that diversity at a topic level is e ective for capturing this e ect in blogs, in news articles, and on Twitter. The phenomenon is present in three distinctly di erent media types, each with their own unique features. We describe the phenomenon using case studies relating to major news stories from September 2015.
The problem of detecting breaking news events has inspired a host of approaches, extracting useful signals from activity on social networks, newswire, and other types of media. The online communication platforms that have been adopted allow these events to persist in some form. These digital traces can never fully capture the original experience, but o er us an opportunity to revisit signi cant phenomena with different points of view, or help us to characterise and learn something about the processes involved. Many di erent forms of news media attempt to record and disseminate information deemed important enough to communicate, and as the barriers to broadcasting and sharing information are removed, attention becomes a scarce commodity.
We de ne the problem of detecting attention
dominating moments across di erent media types, as a
collapse in diversity in the content generated by a set
of online sources in a topic during a given time
period. Media types here include mainstream news
articles, blog posts, and tweets. These media types di er
in both the category of topics covered [
In Section 3, we describe the two stages of our proposed event detection procedure. In the rst stage, content generated by the news, blog and tweet sources is grouped into broad topical categories, through the application of matrix factorization to the content generated by these sources. In the second stage, we examine the variation in similarity between content generated by sources within a given topic during a given time period, in order to identify a collapse in diversity within a topic which corresponds to an attention dominating moment. In Section 5, we evaluate this procedure on a collection of one million news articles and blog posts from September 2015, along with a parallel corpus of tweets collected during the same time period.
Rather than formulating the problem as tracking
the evolution of topics themselves, we consider the
diversity of content within a speci c topic over time. The
motivation is that, for instance, a collapse in diversity
around a major sporting event will be strongly
evident in certain news sources, but not evident in others.
The distinction is important, as this approach is more
suited to retrospective analysis, when the entire
collection of documents of interest is available. The topics
do not change over time, as opposed to a real-time
setting where topics must be updated as new documents
arrive [
Our main contributions here are: 1) a diversitybased approach of detecting attention dominating news events; 2) a comparison between traditional news sources, blogs, and Twitter during these events. 3) a parallel corpus of newsworthy tweets for the NewsIR dataset. 2
In previous work, attention dominating news
stories have been described as media explosions [
Topic modeling applied to parallel corpora of news
and tweets has been previously explored by a number
of researchers [
How o ine phenomena link to bursty behaviour
online is discussed in [
As a broad overview, the common components involved in detecting high impact, attention dominating news stories include: selecting relevant subsets of documents; representation and feature extraction; constructing time series from features; event detection and analysis. In this paper we concentrate on a single key feature of breaking news: a collapse in content diversity within a xed time window. 3
Our objective is to detect when multiple articles in a topical stream become less diverse, signalling the emergence of an attention dominating news story. We consider attention to a phenomenon as the main driving force behind the decision to produce or broadcast a communication. Using the diversity of content within a time window, we attempt to characterise instances where a particular piece of information becomes dominant. Concretely, for each type of media, NMF is used to assign topics to documents; for documents in a topic, we calculate diversity between documents in a time window. This type of analysis allows us to examine the extent to which the onset of an important breaking news event is accompanied by a collapse in textual content diversity, both within a group of news sources and across di erent media types. 3.1
We apply a Non-negative Matrix Factorization (NMF) topic modeling approach to extract potentially interesting topics from a stream of tweets or set of articles. For each media source, we build a tf-idf weighted termdocument matrix and use this as input to NMF.
We also considered LDA to infer topics in these
datasets. The choice of NMF over LDA was primarily
due to computation time. LDA was signi cantly more
computationally expensive than NMF with NNDSVD
[
The same tf-idf representation used for topic modeling is used in diversity calculations. Each article, blog post or tweet is a tf-idf vector. A separate documentterm matrix is built for each media type. Stopwords and words occurring in fewer than 10 documents are removed.
To measure diversity, we calculate the mean cosine similarity between all unique pairs of articles within a topic for a xed time window. Given a set of documents D in a time window, the diversity is: diversity(D) =
Pi;j2D;i6=j cosSim(Di; Dj )
PjiD=1j 1 i
Where cosSim(Di; Dj ) is the cosine similarity of tf-idf vectors of documents i and j in a time window. In practice, calculating similarities between all pairs of documents can be e ciently performed in parallel, and can be calculated in a matter of seconds.
Longer time windows consider more document pairs, which naturally result in smoother trends. In contrast, shorter time windows are more sensitive to brief attention dominating events, but also false positive spikes|where a small number of articles happen to be similar in content, but do not constitute an attention dominating story.
An alternative to content diversity is also considered. Ignoring document content, and just considering the sources of articles, diversity is calculated with Shannon's Diversity Index:
H0 =
R X pi ln pi i=1
Where pi is the proportion of documents produced by the ith source in a time window of interest, R is total number of sources in a given media type.
Both diversity measures produce a single diversity
value per time window, generating a univariate time
series. Changes in diversity that are 2 standard
deviations away from the mean are naively considered to
be important enough to warrant attention. Exploring
more robust and well established methods for change
point detection such as [
For the case studies described in Section 5, the window length was set to 8 hours. While the fast-paced \24/7 news cycle" is described as a constant ood of information, we nd that all three mediums largely follow a more traditional publishing cycle, with prominent spikes in number of published articles on weekday mornings, and low numbers of articles published outside of normal o ce hours. A more detailed analysis of publishing times and characteristics will be explored in future work. 4
To explore attention dominating news stories, we
apply the method described above to three media
sources: mainstream news, blogs, and tweets. For the
rst two sources, the NewsIR dataset1 is used. For
the nal source, we use our own parallel corpus
collected from Twitter2. In contrast to previous work
[
Gathering all members of such lists covering di
erent countries and topics follows the expert-digest
strategy from [
News Blogs Tweets
18,948 73,403 30,448
Of the original 1 million articles provided, 15,878 were ltered as non-English4 or outside the date range of interest (i.e. created between 2015-09-01 and 201509-31). Tweet language ltering was performed using meta-data provided in the tweet. 5
In order to compare the same topics across di erent media types, we compare the top 10 terms representing the topics from di erent models. Speci cally, when topics from two di erent models have stronglyoverlapping (using Jaccard similarity) top term lists, this indicates that similar events were discussed in both media types.
Topics in a model that do not have any overlapping terms with topics in other models, suggest that content unique to a platform is prominent. For example: the \live, periscope, follow, stream, updates" topic in the tweet corpus has no equivalent among the news or blog topics. This re ects the fact that the Periscope app became popular with journalists for broadcasting short live video streams and Twitter is the main platform where these streams are announced. The \music, album, song, video, band " topic is prominent in the blogs and Twitter, but is not present in news. This may re ect the fact that most Twitter accounts and blogs are far more personal in nature.
An indicative, but not necessary feature of attention domination news is the presence of a similar topic on multiple platforms. To illustrate the phenomenon of topical diversity collapse, we now describe three case studies.
4https://github.com/optimaize/language-detector was used for language detection. Interestingly, language detection proved e ective for ltering \spammy" articles containing obfuscated text, large numbers of urls, or containing tabular data.
For each case study, we present the following: Top 10 topic terms for a topic in a media type, and a plot of diversity over time, where:
Solid lines show diversity of documents over time. Dashed lines show Shannon Diversity of sources. Highlighted time periods are when major developments occurred|based on Wikipedia Current Events Portal5 for September 2015.
Dot and Triangle markers indicate periods when diversity drops 2 standard deviations below the mean. 5.1
The European crisis began in 2015, as increasing
numbers of refugees from areas in Syria, Afghanistan, and
Western Balkans [
Blogs
The downward trend in diversity between September 3rd and 5th in the refugee crisis topic can be explained by the death of Aylan Kurdi. News of his 5https://en.wikipedia.org/wiki/Portal:Current_ events/September_2015 Blogs News drowning quickly spread online and made global headlines. This was a particularly far-reaching story, dominating news coverage until an announcement on relaxing controls on the Austro-Hungarian border by Chancellors Faymann of Austria and Merkel of Germany. Both Twitter and mainstream news streams experienced a diversity collapse, while Blogs maintained more diverse set of articles. Between 19th and 21st, smaller drops in diversity coinside with Pope Francis' visit, where the issue of refugees was a prominent topic of discussion. 5.2
Donald Trump's presidential campaign has attracted considerable attention across all types of media6. Positions on issues of immigration and religion are particularly polarising, frequently causing controversies in mainstream media.
Blogs News Tweets
Top 10 Topic Terms trump, donald, republican, presidential, debate, gop, president, candidates, candidate, bush trump, republican, presidential, donald, debate, clinton, bush, orina, candidates, campaign trump, im, love, donald, going, debate, happy, gop, president, think 1.0 0.0 1.0 2.0 3.0 4.0 5.0 1.0 0.0 1.0 2.0 3.0 4.0
Blogs
News 1.0 0.0 1.0 2.0 3.0 4.0 65..00 Tweets 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Figure 2: Standardised diversity scores for Donald Trump Presidential Campaign topic
Signi cant events marked around 12th, 17th, 21st in Figure 2 relate to: Trump's comments on Senator Rand Paul on Twitter which was discussed on mainstream news around 12th, but not as prominently on blogs. On the 16th-17th coverage of a republican presidential debate hosted by CNN; and 21st|mainstream news coverage of reactions to events on 17th: during 6https://en.wikipedia.org/wiki/Donald_Trump_ presidential_campaign,_2016 a town hall meeting in Rochester, Donald Trump declined to correct a man who said that President Obama is a Muslim.
The statement prompted a signi cant drop in the diversity of stories across all platforms. On the 25th, during a speech given to conservative voters in Washington, Trump called fellow Republican presidential candidate Marco Rubio \a clown". Based on the data, it appears that the reaction to the latter on Twitter was not as pronounced as among journalists and bloggers. 5.3
The visit of Pope Francis spanned 19 to 27 September 2015, where the itinerary included venues in both Cuba and the United States. This event is a good illustrative example as it was widely documented7, and highlights a case where a collapse in diversity did not occur at the same time on di erent media platforms.
Blogs
In the case of news publishers, the largest drop in diversity coincided with the beginning of the Pope's visit to Havana. Twitter users and bloggers reacted more on September 23rd and 24th, when the Pope met with Barack Obama and became the rst Pope to address a joint session of US Congress.
7https://en.wikipedia.org/wiki/Pope_Francis'_2015_ visit_to_North_America
In the Twitter stream, the notable event around 16th-17th is due to large numbers of similar tweets as preparations for the visit were being discussed, and #TellThePope trended brie y.
Earlier in the month, we see evidence of overlapping attention dominating events. Between 6th and 7th September, the Pope announced the Vatican's churches will welcome families of refugees. This announcement followed a signi cant development in the ongoing European refugee crisis: around 6,500 refugees arrived in Vienna following Austria's and Germany's decision to waive asylum system rules. This suggests that an attention dominating news event in one topic can trigger events in other topics, especially where prominent public gures are involved. 6
While the diversity measure we propose is relatively
simple, it can be easily augmented to account for other
factors. In the simplest form, every similarity value
between a unique pair of articles within a time window
carries an equal weight in the diversity calculation,
implying that a strong similarity between two highly
in uential publishers is just as important as between
two inconsequential publishers with a small audience.
However, this weight could be tuned, either manually
or automatically using external information (e.g. Alexa
rankings). Accounting for social context [
The sequence of events in the European refugee
crisis and papal visit case studies suggest that it may be
possible to identify and track major developments with
global impact by linking attention dominating
moments across multiple topics, as well as across sources
on di erent platforms. Social media communities both
in uence and are in uenced by traditional news media
[
Acknowledgments: This publication has emanated from research conducted with the support of Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289.