Janette Lehmann, Carlos Castillo, Mounia Lalmas, Ethan Zuckerman
Finding News Curators in Twitter

Outline
¨ Motivation
¨ Types of curators
¨ Labeling news story curators
¨ Automatically finding news story curators
¨ Conclusion and future work
2
Photo credit (first slide): Hobvias Sudoneighm (CC-BY).

Motivation
¨ Twitter has become a powerful tool for the aggregation and consumption of time-
sensitive content in general and news in particular.
¨ Journalists use online social media platforms (Twitter, Facebook and others) and
blogs to elicit other story angles or verify stories they are working on.
To what extend the community of engaged readers - those who
share news articles in social media – can contribute to the journalistic process?
What kind of roles people play when sharing news?
We want to detect users that provide further relevant
information to a news story. We call them news story curators.
3

Example
Al Jazeera English news article about the civil war in Syria
“Syria allows UN to step up food aid” [16 Jan 2013]
Users that posted the article in Twitter
Whom would you follow to find out more about the civil war in Syria?
4
#Followers Is tweeting about
@RevolutionSyria 88,122 Syria
@KenanFreeSyria 13,388 Syria
@UP_food 703 Food
@KeriJSmith 8,838 Breaking news/top stories
@BreakingNews 5,662,866 Breaking news/top stories

Types of news story curators
Human Automatic
Topic-
unfocused
Topic-unfocused curator
Disseminating news articles about
diverse topics, usually breaking
news/top stories
à @KeriJSmith
News aggregators
Collecting news articles (e.g. from
RSS feeds) and automatically post
their corresponding headlines and
URLs
à @BreakingNews
Topic-
focused
Topic-focused curator
Collecting interesting information
with a specific focus, usually a
geographic region or a topic
à @KenanFreeSyria
Topic-focused aggregators
Disseminating automatically news
with topical focus
à @UP_food, @RevolutionSyria
5

Types of news story curators
Human Automatic
Topic-
unfocused
Topic-
focused
Topic-focused curator
Collecting interesting information
with a specific focus, usually a
geographic region or a topic
à @KenanFreeSyria
Topic-focused aggregators
Disseminating automatically news
with topical focus
à @UP_food, @RevolutionSyria
Valuable curators for
a specific story
These curators are probably
less or not valuable
6

Data sets
Step 1: Selection of news articles
¨ News articles published in early 2013 from
¤ BBC World Service [BBC] 75 articles
¤ Al Jazeera English [AJE] 155 articles
¨ Stories: Obama's inauguration, Mali conflict, Pollution in Beijing, etc.
Step 2 : News crowd detection
¨ All users who tweeted the article within the first 6 hours after
publication
Step 3: User characteristics
¨ Extraction of data from each user in the news crowd (e.g. further
tweets, profile information)
7

Labeling
News Story Curators
8
Photocredit:ThomasLeuthard(CCBY).

Labeling tasks
Data
¨ Sample of 20 news articles
¨ For each news article, a sample of 10 users who posted the article
¨ We shown to three assessors:
¤ The title of the news article and a sample of tweets of the user
¤ Profile description and the number of followers of the user
Labeling-Questions
9
Q1) Please indicate whether the user is interested or
an expert of the topic of the article story:
Yes: Most of her/his tweets relate to the topic of the story (e.g.
the article is about the conflict in Syria, she/he is often tweeting
about the conflict in Syria).
Maybe: Many of her/his tweets relate to the topic of the story or
she/he is interested in a related topic (e.g. the article is about the
conflict in Syria, she/he is tweeting about armed conflicts or the
Arabic world).
No: She/he is not tweeting about the topic of the story.
Unknown: Based on the information of the user it was not
possible to label her/him.
Q2) Please indicate whether the user is a human or
generates tweets automatically:
Human: The user has conversations and personal comments in his
or her tweets. The text of tweets that have URLs (e.g. to news
articles) seems self-written and contain user own opinions.
Maybe automatic: The Twitter user has characteristics of an
automatic profile, but she/he could be human as well.
Automatic: The tweet stream of the user looks automatically
generated. The tweets contain only headlines and URLs of news
articles.
Unknown: Based on the information of the user it was not
possible to label her/him as human or automatic.

Resulting training set
Interested?
(topic-focused)
Human or Automatic? Interested
+ human
n yes no n human automatic
AJE 63 21% 79% 71 55% 45% 13%
BBC 58 3% 54% 54 35% 65% 1.8%
many users are
topic-unfocused and automatic
10
We considered only users for which at least two annotators provided a decisive label
(Yes or No, Human or Automatic)

Automatically
finding
News Story Curators
11
Photocredit:MadsIversen(CCBY-NC-SA).

Features
Visibility
• Number of followers
• Number of Twitter lists with user
Tweeting activity
• Number of tweets per day
• Fraction of tweets that contains a re-tweet mark "RT", a URL, a user
mention or a hashtag
Topic focus
• Number of crowds the user belongs to
• Number of distinct article sections of the crowds (e.g. sports, business) the
user belongs to
12

Simple models
UserIsHuman
UserFracURL >= 0.85
automatic,
otherwise human
Model
Human class:
Prec/Rec: 0.85
AUC: 0.81
Evaluation
UserIsInterestedInStory
UserSectionsQ >= 0.9
not-interested,
otherwise interested
Model
Interested class:
Prec: 0.48 / Rec: 0.93
AUC: 0.83
Evaluation
Preselection
The user must have
• At least 1,000 followers
• Posted an article that is estimated related to the original article [1]
13
[1] J. Lehmann, C. Castillo, M. Lalmas, and E. Zuckerman. Transient news crowds in social media. In ICWSM, 2013.
feature (one) selection + random forest algorithm

Complex models
Precision Recall AUC
Automatic 0.88 0.84 0.93
Human 0.82 0.86 0.93
Interested 0.95 0.92 0.90
Not-interested 0.53 0.67 0.90
random forest with
information-gain-based
feature selection
random forest with
asymmetric misclassification costs
false negatives (classifying an interested user
as not interested) were considered 5 times more
costly than false positives
14

Precision-oriented evaluation
We compared our method with two baseline approaches
¨ Users with the largest number of followers [FOLLOWER-APPROACH]
¨ Users with the largest number of stories detected as related to the original one [STORY-APPROACH]
Data
¨ Sample of 20 news articles that had at least one curator, detected using the complex model
with a confidence value >= 0.75
¨ We extracted for each article the same number of possible curators using the other two
approaches
¨ We asked three assessors to evaluate the results
(question Q1 – UserIsInterestedInStory)
¨ About 210 labels for 70 units were collected
Results
true positive/false positive
FOLLOWER-APPROACH: 2/18 = 11%
STORY-APPROACH: 5/20 = 25%
OUR APPROACH: 6/16 = 38%
15

Conclusion and future work
We were able to detect and model news story curators, who (could and maybe are)
play an important role in the news ecosystem; not only for news readers,
but for journalists and editors.
¨ A large amount of activity on Twitter is automatic and some of these news
aggregators can be considered to be good curators
¨ Mostly the attention of the user is quickly shifting away - posting a link does not
have to reflect a long-standing interest on the subject of the link
Future work
¨ Adding other (Twitter) variables to the system that capture, for instance,
interestingness and serendipity
¨ Application on other news providers
¨ Analysis of the functionality of popular news aggregators, which are comparable to
RSS feeds
16

Questions and Discussion…
17
Janette Lehmann
Universitat Pompeu Fabra
jnt.lehmann@gmail.com
Carlos Castillo
Qatar Computing Research
Institute
chato@acm.org
Mounia Lalmas
Yahoo! Labs
mounia@acm.org
Ethan Zuckerman
MIT Center for Civic Media
ethanz@media.mit.edu
Photocredit:WayneLarge(CC-BY-ND).
Photo credits: Hobvias Sudoneighm (CC BY), Thomas Leuthard (CC BY), Mads Iversen (CC BY-NC-SA), Wayne Large (CC BY-ND)