Conference PaperPDF Available

Headlines Matter: Using Headlines to Predict the Popularity of News Articles on Twitter and Facebook



Social media like Facebook or Twitter have become an entry point to news for many readers. In that scenario, the headline is the most prominent – and often the only visible – part of the news article. We propose a novel task of using only headlines to predict the popularity of news articles. The prediction model is evaluated on headlines from two major broadsheet news outlets – The Guardian and New York Times. We significantly improve over several baselines, noting differences in the model performance between Facebook and Twitter.
Headlines Matter: Using Headlines to Predict the Popularity of News Articles on
Twitter and Facebook
Alicja Piotrkowicz
Vania Dimitrova
School of Computing
University of Leeds, UK
Jahna Otterbacher
Social Information Systems
Open University of Cyprus
Katja Markert
Institut f¨
ur Computerlinguistik
at Heidelberg, Germany
Social media like Facebook or Twitter have become an entry
point to news for many readers. In that scenario, the headline
is the most prominent – and often the only visible – part of
the news article. We propose a novel task of using only head-
lines to predict the popularity of news articles. The prediction
model is evaluated on headlines from two major broadsheet
news outlets – The Guardian and New York Times. We signif-
icantly improve over several baselines, noting differences in
the model performance between Facebook and Twitter.
Introduction and Related Work
Headlines are vital in both capturing readers’ attention and
in influencing their online reading experience of news. In
fact, approximately six in 10 people limit their reading to
headlines only, without clicking on a link to the full article1.
Eye-tracking studies have confirmed this behaviour empiri-
cally; many people are “entry-point readers”, who attend to
headlines in order to ascertain the overview of an article, but
who exhibit minimal reading activities (Holsanova, Rahm,
and Holmqvist 2006). Furthermore, there are many online
spaces where headlines are the only visible part of the news
article; for example news feeds and social media.
Yet despite this, headlines have not been considered be-
fore as the sole source of data for news article popular-
ity prediction. Most models make use of post-publication
data, such as the number of early adopters (Castillo et al.
2014). These methods model popularity development, e.g.
they might use the number of tweets within the first hour
after article publication to predict later or final popularity.
On the other hand, approaches which tackle what Ara-
pakis, Cambazoglu, and Lalmas (2014) call the “cold start
problem”, i.e. the prediction of news article popularity prior
to publication, are still in their infancy. In particular, these
approaches offer limited insight into which aspects of the
news article text make it popular online. Bandari, Asur, and
Huberman (2012) use a small number of text features re-
lated to topical category, named entities’ prominence and
sentiment. Arapakis, Cambazoglu, and Lalmas (2014) re-
produce and improve upon the work by Bandari, Asur, and
Copyright c
2017, Association for the Advancement of Artificial
Intelligence ( All rights reserved.
Huberman (2012). They also add a small number of linguis-
tic and prominence features, but their main focus is on eval-
uation methods. Both Bandari, Asur, and Huberman (2012)
and Arapakis, Cambazoglu, and Lalmas (2014) use the news
source as a feature, which is shown to be the overwhelm-
ing determiner of popularity. However, if the newsroom staff
want to adjust article content to reach larger audiences, this
is unhelpful, as news source is out of their control. More-
over, these previous models consider headlines and article
body jointly. As headlines play a crucial role in the online
news domain, it is worth investigating to what extent we can
predict an article’s popularity from the headline alone. Our
goal is to investigate a wide variety of text features extracted
from headlines and determine whether they have impact on
social media popularity of news articles. We enhance prior
work by: (i) using only headlines; (ii) introducing new fea-
tures; and (iii) using a source-internal evaluation.
Data Collection
We created two corpora of news headlines and obtained the
social media popularity for each headline.
News corpora. We used two major broadsheet newspa-
pers — The Guardian and New York Times. We downloaded
all headlines published during April 2014 (Guardian train-
ing), July 2014 (Guardian test), October 2014 (NYT train-
ing), and December 2014 (NYT test)2. Table 1 includes
some example headlines with their popularity scores.
Social media data. We measure a news article’s social
media popularity by the number of times it is cited on Twit-
ter and Facebook. The article URL was used as the search
query for the Twitter Search API3to obtain the number of
tweets and retweets one, three, and seven days after the ar-
ticle’s publication. The process was repeated for Facebook
likes and shares using the Facebook FQL API.4
Popularity measures. Tweets and retweets, as well as
shares and likes, are combined into two metrics: Twitter and
Facebook popularity. We found that in our datasets Twitter
and Facebook popularity after three and seven days did not
differ significantly, and so throughout the paper we report
2The Guardian data: Guardian Content API, New York Times:
NYT Article Search API.
Table 1: Examples of the most and least popular headlines.
The Guardian New York Times
Most popular “Capitalism simply isn’t working and here are the reasons
why” (T=2299, F=23840)
“Doctor in New York City Is Sick With Ebola” (T=12780,
Least popular “Corrections and clarifications” (T=0, F=0) “Pastis and Ouzo: The Soccer of Liquors” (T=0, F=0)
“Fears misplaced over letting Lords resign” (T=5, F=0) “Alaska’s Political Outlook” (T=0, F=1)
popularity after three days, yielding two social media popu-
larity measures: T = Twitter popularity after three days, and
F = Facebook popularity after three days.
Data overview. The popularity measures show a strongly
Zipfian distribution. Twitter and Facebook measures corre-
late well with each other (Guardian: ρ=0.74, NYT: ρ=0.6).
However, Twitter shows a flatter distribution than Facebook.
In both datasets the number of citations is much higher for
Facebook rather than Twitter, which might be due to the
number of users (in 2016 Facebook had 1.7 billion active
users to Twitter’s 0.3 billion5). News source also plays an
important role, as New York Times articles are more often
shared on social media (this follows the finding by Bandari,
Asur, and Huberman (2012) that news source is the strongest
predictor of social media popularity of news articles).
Headline Features
We use two types of features: journalism-inspired news val-
ues and linguistic style. Details of feature implementations
are outlined in Piotrkowicz, Dimitrova, and Markert (2017).
News Values
News values, a concept originating in journalism studies, re-
fer to aspects of news stories which make them newsworthy.
While there are many news values taxonomies, there is con-
siderable overlap (cf. Caple and Bednarek (2013)) and we
implement six news values which are frequently included.
Prominence. Reference to prominent entities is one of the
key news values. We approximate Prominence as the amount
of online attention an entity gets. We extend previous work
by using wikification for obtaining entities, which ensures
a wide variety of entity types. We implement six Promi-
nence features: (i) number of wikified entities; (ii) news re-
cent prominence (the number of entity mentions in head-
lines of the relevant news outlet); (iii) long-term prominence
(the median number of daily Wikipedia page views over a
year for an entity); (iv) day-before prominence (number of
Wikipedia page views for an entity the day before a given
headline was published); (v) current burst size (if an entity
is ‘bursty’, how much above the average are the entity’s page
views); and (vi) burstiness (if an entity is ‘bursty’, how many
times over a year the entity has been in a burst).
Sentiment. Sentiment refers to both negative and pos-
itive vocabulary. We calculate four features using Senti-
WordNet (Baccianella, Esuli, and Sebastiani 2010) scores:
(i) sentiment (maxP os maxNeg 2); (ii) polarity
(maxP os+maxN eg); (iii) proportion of biased words; and
(iv) proportion of positively/negatively connotated words.
Magnitude. This refers to the size or impact of a news
event. There are three features: (i) proportion of comparative
and superlative words (based on POS tags); (ii) proportion of
intensifiers; and (iii) proportion of downtoners.
Proximity. We focus on geographic proximity to the news
source, which assumes that readers from the same country as
the news outlet constitute a large part of its readership. We
implement Proximity as an explicit reference to news outlet
country (UK/US related keywords) in the headline text.
Surprise. Surprising headlines draw attention. We mea-
sure surprise by calculating the commonness of syntactic
chunks in a headline with reference to a Wikipedia corpus6.
Uniqueness. Headlines should be novel. To investigate
this, for a given headline we go through recent past head-
lines to see if there are any highly similar ones. For a pair
of headline and past headline vectors (created using a tf-idf
weighted Gigaword corpus) we calculate cosine similarity.
The highest cosine similarity is assigned as the feature value.
For linguistic style we calculate features inspired by journal-
ism studies and NLP work on the effect of phrasing.
Brevity. Headlines need to be short. We implement
Brevity as the number of tokens and number of characters.
Simplicity. Easy-to-understand headlines use simple syn-
tax and vocabulary. We use two syntactic complexity fea-
tures: (i) parse tree height, and (ii) number of non-terminal
tree nodes. To measure lexical complexity we implement
four features: (i) entropy (calculated using a trigram lan-
guage model built using the CMU-Cambridge Toolkit on the
New York Times section of the Gigaword corpus); (ii) propor-
tion of difficult words (any word not occurring among the
5000 most common words in the language model); (iii) me-
dian word frequency (using the unlemmatised word fre-
quency lists7); and (iv) information content (calculated for
nouns and verbs on British National Corpus).
Unambiguity. News text should not be ambiguous. We
use two features for headline ambiguity: median number of
senses per word from WordNet, and modality (modal event
or modal relation between events; using TARSQI8).
Punctuation. The Guardians style guide for headlines9
7; British National Corpus for
Guardian, Corpus of Contemporary American English for NYT.
discourages using quote, question, and exclamation marks.
We implement binary features indicating their presence.
Nouns. The Guardians style guide cautions against us-
ing too many successive nouns (so-called ‘headlinese’). We
implement four features: (i) three consecutive nouns (bi-
nary feature); (ii) number of noun phrases; (iii) proportion
of common nouns; and (iv) proportion of proper nouns.
Verbs. Using verbs is encouraged in headlines in The
Guardians style guide. We implement two features: (i)
number of verb phrases, and (ii) proportion of verbs.
Adverbs. Adverbs, especially adverbs of manner, are fre-
quently used in headlines. We use the proportion of adverbs.
Prediction Model
Using these features, our goal is to predict the popularity
of news articles on Twitter and Facebook from the article’s
headline. We build separate prediction models for each news
source, thus avoiding news source popularity effects.
We used regression (support vector regression with RBF ker-
nel). Arapakis, Cambazoglu, and Lalmas (2014) argue that
using classification for popularity prediction is not appropri-
ate, as class splits potentially introduce bias towards articles
with low popularity. Popularity measures – T, F – were log-
transformed in order to improve model fit.
Results were evaluated on the test set10. Two evaluation
metrics were used: Kendall’s tau rank correlation coefficient
(τ) and mean absolute error (MAE). Significance testing was
performed using z-test for τand t-test for MAE.
We used three baselines: a unigrams baseline and two state-
of-the-art baselines. Our model’s features are denoted as M.
Unigrams (MU). We used 1000 most frequent unigrams.
State-of-the-art reimplementations: Bandari, Asur, and
Huberman (2012) (MB) and Arapakis, Cambazoglu, and
Lalmas (2014) (MA) originally used full article text, but we
ran these baselines on the same dataset as ours (i.e. head-
lines only). We aimed at as close a reimplementation as pos-
sible, but in some cases we had to make adjustments. We
used Stanford Named Entity Recognizer and SentiWordNet
for Prominence and Sentiment features, respectively. With-
out access to archival Twitter data, we used Wikipedia to
calculate Prominence features. Finally, unlike in the original
implementations there is no news source feature (as our goal
is a source-internal evaluation).
The two state-of-the-art approaches, as well as similar
tasks (Lakkaraju, McAuley, and Leskovec 2013), make use
of metadata that is available at the time of article publication
(category, time). The reimplemented baselines and our full
model (M) also include metadata. Following the implemen-
tation by Arapakis, Cambazoglu, and Lalmas (2014), both
category and publication date and time are implemented as
10Evaluation using cross-validation is not appropriate, because
the data is temporally ordered and one of our features, headline
uniqueness, makes use of the temporal ordering.
binary features in our model. Bandari, Asur, and Huberman
(2012) calculate a category score (#citations per categor y
#articles per categor y ).
Results and Discussion
We report regression results against different baselines.
Table 2: Regression results of baselines against our model
(M) using all features (news values, style, metadata). Result
in bold indicates improvement of p<0.05.
The Guardian New York Times
MU0.32 0.25 0.82 1.59 0.19 0.22 0.66 1.68
MB0.36 0.29 0.71 1.53 0.15 0.18 0.67 1.72
MA0.41 0.35 0.7 1.45 0.21 0.3 0.86 1.57
M0.43 0.37 0.68 1.42 0.23 0.32 0.88 1.54
Table 3: Regression results of baselines against our model
(M) using headline features only (news values and style).
Result in bold indicates improvement of p<0.05.
The Guardian New York Times
MB0.11 0.07 0.94 1.74 0.05 0.02 0.7 1.85
MA0.22 0.19 0.88 1.66 0.19 0.16 0.67 1.75
M0.29 0.26 0.83 1.59 0.21 0.23 0.69 1.66
Performance against baselines using full feature set
(Table 2). Our model (M) significantly outperforms the
baselines for nearly all measures. The exceptions are the
MAE results for Twitter in the NYT dataset, where uni-
grams baseline outperforms the model. However, for that
same dataset our model achieves significantly higher cor-
relations. The best results were achieved for Twitter in the
Guardian dataset (τ=0.43, MAE=0.68). This is a promising
result, considering that this is the first attempt to use head-
lines for a news article popularity prediction task.
Performance against baselines using content features
only (Table 3). When limited to features that can be ex-
tracted directly from headline text (news values and style),
our model shows a considerably better improvement for
most measures than when comparing models with metadata
(improvement in correlation of approx. 40%, compared to
5-10% when using metadata). The highest correlation was
achieved for Twitter in the Guardian dataset (τ= 0.29), and
the lowest MAE for Twitter in NYT dataset (MAE=0.69).
Performance of feature groups (Table 4). Using all fea-
tures significantly outperforms any individual feature group
at p<0.01. Again, the exceptions are MAE results for Twit-
ter in the NYT dataset. Although news values achieve the
Table 4: Regression results comparing feature groups (MN
= news values, MS= style, MM= metadata). Result in
bold indicates improvement of p<0.01.
The Guardian New York Times
MN0.2 0.17 0.89 1.67 0.14 0.14 0.68 1.74
MS0.25 0.22 0.86 1.62 0.18 0.19 0.7 1.7
MM0.39 0.33 0.72 1.51 0.17 0.23 0.92 1.65
M0.43 0.37 0.68 1.42 0.23 0.32 0.88 1.54
lowest performance of all groups, the correlation with Twit-
ter and Facebook popularity is still between 0.14 and 0.2.
It is especially noteworthy that style features, which are
largely topic-independent, on their own achieve good per-
formance (up to 0.25 correlation, and 0.7 MAE). This sug-
gests that headline style is important to social media readers,
independent of article content. This seems to follow previ-
ous research on online content popularity prediction, where
various aspects of style were also found to have an impact
on popularity (Tan, Lee, and Pang 2014). Metadata (espe-
cially category) achieves good results, in particular for the
Guardian dataset, suggesting that topic and genre of the ar-
ticle play a significant role for readers. Although metadata
adds to the prediction performance, it should be noted that
this aspect of news articles is usually not controlled by the
writer (i.e. one cannot easily change the genre or the topic).
On the other hand, most news values and style features can
be freely edited, in order to reach higher popularity.
Differences between Twitter and Facebook. Perhaps
due to a more skewed distribution, Facebook has higher
errors. For correlations, Twitter performs higher with
Guardian data, while it is the opposite for NYT. Different
demographics of news readers on Facebook and Twitter11
might contribute to this, which calls for further work that
takes into account user demographics.
Differences between news sources. A key aspect of our
work is the source-internal evaluation, which has not been
done for this task before. Indeed, the performance is better
on Guardian data than the NYT. This points at further work
with other news outlets and genres (e.g. tabloids).
Computation cost vs. performance. When comparing
against the best-performing baseline (MA), our full model
achieves significant improvement (cf. Table 2). The differ-
ence is much more noticeable when considering only fea-
tures extracted directly from headline text (cf. Table 3).
While the overall performance using all available features
only slightly improves over the state-of-the-art, the model
that uses features that can be more readily edited by the
headline author (and possibly increase its popularity) shows
considerable improvement.
News headlines play a crucial role on social media. In a
novel task to predict the social media popularity of news ar-
ticles using headline-derived features, we improved signifi-
cantly over several baselines. Features extracted from head-
line text (which usually can be edited by the headline author)
were shown to have impact on the prediction performance
when considered on their own. This suggests that traditional
editorial judgments about newsworthiness and insights from
NLP research on style are applicable to predicting headline
popularity on social media. Our feature extraction methods
are generic and can be repeated across different news out-
lets and genres. The results of the prediction model depend
on the news source; further work can include performance
comparison across different news outlets and online content.
We are currently refining the prediction model taking into
account user demographics and integrating world knowl-
edge. Firstly, we are considering user location (country of
residence) to improve the Proximity feature. Secondly, to
improve Prominence (our best-correlated feature) we are in-
corporating world knowledge from Wikidata to relate entity
significance to the user’s location.
This work was supported by a Doctoral Training Grant from
the EPSRC, UK. Data collection and storage comply with
EPSRC data management policies. The dataset is available
Arapakis, I.; Cambazoglu, B. B.; and Lalmas, M. 2014. On
the feasibility of predicting news popularity at cold start. In
SocInfo, 290–299.
Baccianella, S.; Esuli, A.; and Sebastiani, F. 2010. Sen-
tiwordnet 3.0: An enhanced lexical resource for sentiment
analysis and opinion mining. In LREC.
Bandari, R.; Asur, S.; and Huberman, B. A. 2012. The pulse
of news in social media: Forecasting popularity. In ICWSM.
Caple, H., and Bednarek, M. 2013. Delving into the dis-
course: Approaches to news values in journalism studies and
beyond. Reuters Institute for the Study of Journalism.
Castillo, C.; El-Haddad, M.; Pfeffer, J.; and Stempeck, M.
2014. Characterizing the life cycle of online news stories
using social media reactions. In CSCW.
Holsanova, J.; Rahm, H.; and Holmqvist, K. 2006. Entry
points and reading paths on newspaper spreads: comparing
a semiotic analysis with eye-tracking measurements. Visual
communication 5(1):65–93.
Lakkaraju, H.; McAuley, J. J.; and Leskovec, J. 2013.
What’s in a Name? Understanding the Interplay between Ti-
tles, Content, and Communities in Social Media. In ICWSM.
Piotrkowicz, A.; Dimitrova, V. G.; and Markert, K. 2017.
Automatic extraction of news values from headline text. In
Proceedings of the EACL 2017 Student Research Workshop.
Tan, C.; Lee, L.; and Pang, B. 2014. The effect of wording
on message propagation: Topic-and author-controlled natu-
ral experiments on Twitter. In ACL.
... Time-series based solutions handle the forwarding prediction problem by feature based methods, temporal process methods and deep learning methods. Feature based methods extract hand-crafted features (Piotrkowicz et al. 2017;Shulman, Sharma, and Cosley 2016;Keneshloo et al. 2016) and construct machine learning models for prediction. A number of models were proposed to describe the evolution of accumulation process of information diffusion volumes (Shen et al. 2014;Cui et al. 2013;Gao, Ma, and Chen 2015;Mishra, Rizoiu, and Xie 2016;Rizoiu et al. 2017). ...
... Please notice that this paper focuses on cascade graph modeling, some existing popularity prediction methods, complemented by other features such as text content (Liao et al. 2019;Piotrkowicz et al. 2017), publishers and users' personal information (Chen et al. 2019a;Zhang et al. 2018), are complementary directions and can be plugged into Tem-pCas. Thus, they are not included in our baselines. ...
Real-time forwarding prediction for predicting online contents' popularity is beneficial to various social applications for enhancing interactive social behaviors. Cascade graphs, formed by online contents' propagation, play a vital role in real-time forwarding prediction. Existing cascade graph modeling methods are inadequate to embed cascade graphs that have hub structures and deep cascade paths, or they fail to handle the short-term outbreak of forwarding amount. To this end, we propose a novel real-time forwarding prediction method that includes an effective approach for cascade graph embedding and a short-term variation sensitive method for time-series modeling, making the best of cascade graph features. Using two real world datasets, we demonstrate the significant superiority of the proposed method compared with the state-of-the-art. Our experiments also reveal interesting implications hidden in the performance differences between cascade graph embedding and time-series modeling.
... Popularity and reliability of news articles shared on social media have been studied before as separate topics. Efforts on predicting popularity of news articles often rely on hand-crafted content features [1,2,34] and early engagement statistics [4,27,56]. Prior work focuses on textual content, and does not investigate in what way accompanying visuals contribute to popularity, even though modern media is often multi-modal. ...
... For example, [2] use textual features such as topic, sentiment and named entities mentioned in the article. [34] shows article titles reveal strong signals for popularity but its dataset is limited to two news sources. Some recent work investigates popularity of a specific type of content such as images [9,61,63] and videos [3,17,50]. ...
Social media content routinely incorporates multi-modal design to covey information and shape meanings, and sway interpretations toward desirable implications, but the choices and outcomes of using both texts and visual images have not been sufficiently studied. This work proposes a computational approach to analyze the outcome of persuasive information in multi-modal content, focusing on two aspects, popularity and reliability, in COVID-19-related news articles shared on Twitter. The two aspects are intertwined in the spread of misinformation: for example, an unreliable article that aims to misinform has to attain some popularity. This work has several contributions. First, we propose a multi-modal (image and text) approach to effectively identify popularity and reliability of information sources simultaneously. Second, we identify textual and visual elements that are predictive to information popularity and reliability. Third, by modeling cross-modal relations and similarity, we are able to uncover how unreliable articles construct multi-modal meaning in a distorted, biased fashion. Our work demonstrates how to use multi-modal analysis for understanding influential content and has implications to social media literacy and engagement.
... Many studies have eventuated previously in this area, as many online news portals are ascending to get the idea of how the news gain popularity. News popularity prediction has been studied in various target domains of social media including, Facebook, and Twitter, [1]- [4], and in different languages [5]- [9], where, prediction can be performed either after or before publication. Most of the earlier works focus on news popularity prediction after publication. ...
Full-text available
The development of world wide web with easy access to massive information sources anywhere and anytime paves way for more people to rely on online news media rather than print media. The scenario expedites rapid growth of online news industries and leads to substantial competitive pressure. In this work, we propose a set of hybrid features for online news popularity prediction before publication. Two categories of features extracted from news articles, the first being conventional features comprising metadata, temporal, contextual, and embedding vector features, and the second being enhanced features comprising readability, emotion, and psycholinguistics features are extracted from the articles. Apart from analyzing the effectiveness of conventional and enhanced features, we combine these features to come up with a set of hybrid features. We curate an Indian news dataset consisting of news articles from the most rated Indian news websites for the study and also contribute the dataset for future research. Evaluations are performed over the Indian news dataset (IND) and compared with the performance over the benchmark mashable dataset using various supervised machine learning models. Our results indicate that the proposed hybrid of enhanced features with conventional features are highly effective for online news popularity prediction before publication.
... Studies in other online communities have also found that title plays a critical role in content popularity. On a dataset of online news articles, Piotrkowicz et al. (2017) improve on popularity prediction baselines using a title-only model and propose that the content title is often "the most prominent" component of online content. Undoubtedly, content titles serve as useful hooks to draw initial attention. ...
Full-text available
Several factors influence the popularity of content on social media, including the what, when, and who of a post. Of these factors, the what and when of content are easiest to customize in order to maximize viewership and reach. Further, the title of a post (part of the what) is the easiest to tailor, compared to the post's body, which is often fixed. So, in this paper, we assess the impact of a post's title on its popularity while controlling for the time of posting (the when) by proposing an interpretable attention-based model. Our approach achieves state-of-the-art performance in predicting the popularity of posts on multiple online communities of various sizes, topics, and formats while still being parsimonious. Interpretation of our model's attention weights sheds light on the heterogeneous patterns in how the specific words in a post's title shape its popularity across different communities. Our results highlight the power of sentiment alignment, personal storytelling, and even personality politics in propelling content to virality.
... Furthermore, the quick contextual cues associated with social media messages coupled with a public that has limited direct experience with certain scientific topics, such as climate change, may influence a lack of trust in science communication (Abrams et al., 2015). Moreover, headlines are often more interesting than the article content leading to audience consumption of half-truths despite the large amounts of in-depth information available (Piotrkowicz et al., 2017). Twitter is known to be effective in facilitating information flow and triggering changes in standard informational structures, as it provides a platform for collaboration and coordination among activist groups (Segerberg & Bennett, 2011). ...
... The traditional approach also applies advanced deep learning model to flip the bias of news headlines. Also, several datasets that present bias or enable the analysis of news and news headline are curated from different sources and discussed in (Horne et al., 2018;Piotrkowicz et al., 2017). ...
Conference Paper
Full-text available
Oncology is one of the most dynamic branches of medicine. As a result of numerous oncology studies, there has been a significant increase in scientific and clinical data that the human brain cannot store. Advances in artificial intelligence (AI) technology have led to its rapid clinical application. In this paper, we wanted to see the role of the use of artificial intelligence (AI) in oncology. We conducted an unsystematic search of databases (Pub Med, MEDLINE, and Google Scholar) using the keywords: artificial intelligence, deep learning, machine learning, oncology, personalized medicine. From a large number of articles available to us, we singled out review articles and clinical trial results according to their clarity and innovation regarding the use of artificial intelligence in oncology. Of particular importance to us was the ability to apply their results in everyday clinical work. The possibilities of using artificial intelligence in oncology are innumerable. Thus, AI can be used for diagnostic purposes (malignant screening, histopathology, and molecular diagnostics), therapeutic purposes (personalized treatment, prediction of treatment side effects and response to therapy, treatment decisions) as well as for prognostic purposes (risk stratification, 5-year survival, monitoring). The implementation of AI in clinical practice presents new challenges for clinicians. Namely, in the era of evidence-based and patient-centered medicine, they will have to master statistical as well as computer skills in addition to clinical ones. Therefore, it is necessary to start educating future doctors about the importance of AI in medicine as soon as possible.
... The traditional approach also applies advanced deep learning model to flip the bias of news headlines. Also, several datasets that present bias or enable the analysis of news and news headline are curated from different sources and discussed in (Horne et al., 2018;Piotrkowicz et al., 2017). ...
Conference Paper
Full-text available
The information on the web can be inconsistent across different web pages. News articles are examples of information on the web that are inconsistent and this paper proposes an approach that enables the visual analysis of inconsistencies in online news. It presents an approach which will enable the visual identification of inconsistencies associated to a news headline of interest. It uses a visual assessment approach that relies on two techniques, namely Fault Tolerance and Co-occurrence techniques. The Fault Tolerance technique is used in extracting related news headlines on the internet while the Co-occurrence technique is used for grouping and scaling related news headlines on the web. Also the bar-chart is used to plot charts that summaries the inconsistencies from which news readers can visually assess related news headlines of particular context.
... On our present focus of news, Berger & Milkman [14] studied the characteristics of New York Times articles that were heavily shared, identifying that articles that express positive or high-arousal emotions have a higher likelihood of becoming popular. A broad literature focuses on predicting success in news by various means [15][16][17][18][19][20][21], although much of this literature prioritizes prediction accuracy above the interpretation of features. The linguistic predictors of success have, however, been studied in other domains. ...
Full-text available
What makes written text appealing? In this registered report protocol, we propose to study the linguistic characteristics of news headline success using a large-scale dataset of field experiments (A/B tests) conducted on the popular website Upworthy comparing multiple headline variants for the same news articles. This unique setup allows us to control for factors that can have crucial confounding effects on headline success. Based on prior literature and a pilot partition of the data, we formulate hypotheses about the linguistic features that are associated with statistically superior headlines. We will test our hypotheses on a much larger partition of the data that will become available after the publication of this registered report protocol. Our results will contribute to resolving competing hypotheses about the linguistic features that affect the success of text and will provide avenues for research into the psychological mechanisms that are activated by those features.
Full-text available
The Guardian editorial headline is viewed as a two-component structure punctuated with colons in which the first part names the topic and the second one provides its comment. The article examines the frequency and diversity of eight noun phrase patterns and gives structural and functional analysis of their constituents. The author studies how categorial features of nouns, adjectives, and prepositions manifest themselves on a phrase level. Three types of semantic relations between noun-noun components are defined. Two more aspects under consideration are complexity and coordination in noun phrases.
Full-text available
This paper presents a study of the life cycle of news articles posted online. We describe the interplay between website visitation patterns and social media reactions to news content. We show that we can use this hybrid observation method to characterize distinct classes of articles. We also find that social media reactions can help predict future visitation patterns early and accurately. We validate our methods using qualitative analysis as well as quantitative analysis on data from a large international news network, for a set of articles generating more than 3,000,000 visits and 200,000 social media reactions. We show that it is possible to model accurately the overall traffic articles will ultimately receive by observing the first ten to twenty minutes of social media reactions. Achieving the same prediction accuracy with visits alone would require to wait for three hours of data. We also describe significant improvements on the accuracy of the early prediction of shelf-life for news stories.
Full-text available
The abstract for this document is available on CSA Illumina.To view the Abstract, click the Abstract button above the document title.
Full-text available
News articles are extremely time sensitive by nature. There is also intense competition among news items to propagate as widely as possible. Hence, the task of predicting the popularity of news items on the social web is both interesting and challenging. Prior research has dealt with predicting eventual online popularity based on early popularity. It is most desirable, however, to predict the popularity of items prior to their release, fostering the possibility of appropriate decision making to modify an article and the manner of its publication. In this paper, we construct a multi-dimensional feature space derived from properties of an article and evaluate the efficacy of these features to serve as predictors of online popularity. We examine both regression and classification algorithms and demonstrate that despite randomness in human behavior, it is possible to predict ranges of popularity on twitter with an overall 84% accuracy. Our study also serves to illustrate the differences between traditionally prominent sources and those immensely popular on the social web.
Conference Paper
Full-text available
In this work we present SENTIWORDNET 3.0, a lexical resource explicitly devised for supporting sentiment classification and opinion mining applications. SENTIWORDNET 3.0 is an improved version of SENTIWORDNET 1.0, a lexical resource publicly available for research purposes, now currently licensed to more than 300 research groups and used in a variety of research projects worldwide. Both SENTIWORDNET ...
Conference Paper
We perform a study on cold-start news popularity prediction using a collection of 13,319 news articles obtained from Yahoo News. We characterise the online popularity of news articles by two different metrics and try to predict them using machine learning techniques. Contrary to a prior work on the same topic, our findings indicate that predicting the news popularity at cold start is a difficult task and the previously published results may be superficial.
Consider a person trying to spread an important message on a social network. He/she can spend hours trying to craft the message. Does it actually matter? While there has been extensive prior work looking into predicting popularity of social-media content, the effect of wording per se has rarely been studied since it is often confounded with the popularity of the author and the topic. To control for these confounding factors, we take advantage of the surprising fact that there are many pairs of tweets containing the same url and written by the same user but employing different wording. Given such pairs, we ask: which version attracts more retweets? This turns out to be a more difficult task than predicting popular topics. Still, humans can answer this question better than chance (but far from perfectly), and the computational methods we develop can do better than both an average human and a strong competing method trained on non-controlled data.
Delving into the discourse: Approaches to news values in journalism studies and beyond. Reuters Institute for the Study of Journalism
  • H Caple
  • M Bednarek
Caple, H., and Bednarek, M. 2013. Delving into the discourse: Approaches to news values in journalism studies and beyond. Reuters Institute for the Study of Journalism.
What's in a Name? Understanding the Interplay between Titles, Content, and Communities in Social Media
  • H Lakkaraju
  • J J Mcauley
  • J Leskovec
  • K Markert
Lakkaraju, H.; McAuley, J. J.; and Leskovec, J. 2013. What's in a Name? Understanding the Interplay between Titles, Content, and Communities in Social Media. In ICWSM. Piotrkowicz, A.; Dimitrova, V. G.; and Markert, K. 2017. Automatic extraction of news values from headline text. In Proceedings of the EACL 2017 Student Research Workshop.