Weeki - Wikipedia <- tweets

UNIVERSITY OF SOUTHAMPTON
FACULTY OF PHYSICAL AND APPLIED SCIENCES
Electronics and Computer Science
Bringing microblog updates to Wikipedia
by
Maël Thomas
A project report submitted for the award of
MSc Web Technology
Supervisor: Dr Leslie Carr
Examiner: Dr Nicholas Gibbins
September 20, 2012

UNIVERSITY OF SOUTHAMPTON
ABSTRACT
FACULTY OF PHYSICAL AND APPLIED SCIENCES
Electronics and Computer Science
A project report submitted for the award of MSc Web Technology
BRINGING MICROBLOG UPDATES TO WIKIPEDIA
by Maël Thomas
Wikipedia is not to be expected to display real-time information. One cannot but
notice that this would indeed be convenient, for its articles are obvious reference
pages for each subject and provide detailed context around that information. We
present Weeki, an application that extracts meaning from Twitter to add a real-
time layer to Wikipedia. This project aims to bridge the gap between these two
Web information giants, through a complex information retrieval step followed
by a careful integration that adds a complementary dimension to both reading
and editing the online encyclopedia. This paper provides a detailed design of the
application features, and a proof of concept implementation.

Contents
Acknowledgements ix
1 Background 3
1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Tweet processing . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1.1 Entity Linking . . . . . . . . . . . . . . . . . . . . 5
1.2.1.2 Information extraction from micro posts . . . . . . 8
1.2.1.3 Network-wide information extraction . . . . . . . . 10
1.2.2 Semantic Microblogging . . . . . . . . . . . . . . . . . . . . 11
1.2.3 Linked Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.4 Web page enrichment . . . . . . . . . . . . . . . . . . . . . . 12
2 The Weeki system 13
2.1 Engine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Two possible approaches . . . . . . . . . . . . . . . . 13
2.1.2 Linking tweets to a Wikipedia article . . . . . . . . . . . . . 14
2.1.3 Stream Enrichment . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Application Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Page Enrichment . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1.1 Smart Tweet Insertion . . . . . . . . . . . . . . . . 17
2.2.1.2 Reference suggestion . . . . . . . . . . . . . . . . . 19
2.2.1.3 Headline Suggestion . . . . . . . . . . . . . . . . . 19
2.2.2 Linked Databases Population . . . . . . . . . . . . . . . . . 19
2.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.1 Software Stack . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2 Extensibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.3 Process description . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.3.1 Anchor Dictionary . . . . . . . . . . . . . . . . . . 23
2.4 Project Management . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Evaluation and future work 25
3.1 Critical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1.1 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
v

vi CONTENTS
3.1.2 Value and limits of our work . . . . . . . . . . . . . . . . . . 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1.3 Reception . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Online Newspapers . . . . . . . . . . . . . . . . . . . 27
Content specialization . . . . . . . . . . . . . . . . . . 27
Personalized streams . . . . . . . . . . . . . . . . . . 28
Wikipedia Redesign . . . . . . . . . . . . . . . . . . . 28
Additional sources . . . . . . . . . . . . . . . . . . . . 28
Generic applicability . . . . . . . . . . . . . . . . . . 28
4 Conclusions 29
References 31

Acknowledgements
I would like to thank my supervisor and second examiner for their help and inter-
esting suggestions. I also thank Raul, Ivan and Daniele for their constant support.
A special thank as well to Camellia Sinensis.
ix

Introduction
Twitter and Wikipedia are two of the most visited websites, both enjoying wide
recognition respectively as a social network and an online and collaborative en-
cyclopedia. While Wikipedia presents semi-structured and high quality content,
Twitter exemplies the real-time Web thanks to its straightforward message model.
We consider them in this work as complementary information sources.
While Wikipedia's information value is now unquestionable
1
, the automated ex-
traction of meaning from Twitter is a complex task, recently illustrated and in-
creasingly subject of attention for the research community.
This project rst focuses on the extraction of information from individual tweets
and from the microblogging network as a whole, their linking to Wikipedia articles
as well as their encoding as semantic data. But more importantly, we give our vi-
sion of a concrete integration of this stream to Wikipedia, designing an application
and providing a technical proof of concept.
Our application is built with a Web page enrichment approach, to assure a cir-
cumspect integration to a website that has built its success on strong guidelines.
We believe that this work provides new and desirable features to the online ency-
clopedia, and moreover draws new possibilities for user contribution. Though not
viable in its current state, we hope that our work will be followed or drive similar
initiatives.
This paper is organised as follows. The background section gives the context, and
introduces relevant natural language tools as well as important related works. The
design of our idea is then presented, followed by a description of our implementa-
tion. We nish with a critical evaluation and indicate interesting future directions
for this project.
1Although of course not perfect and often benchmarked against traditional encyclopedias
1

Chapter 1
Background
1.1 Context
At the center of this project are Wikipedia, the reference online free and collab-
oratively edited encyclopedia,
1
and Twitter, the microblogging platform that has
established itself as a major player on the social Web. Wikipedia's content is well
categorized, and its articles describe concepts of reference for the Web, extracted
to form what can be viewed as a core for the network of Linked Data, DBpe-
dia. Twitter rode the wave of the social web, attracting users and information
sources (blogs, news agencies. . . ) that would make it a full-edged information
medium. It is an immensely vast, varied and above all real-time source of, alas,
unstructured and noisy information, its scope and message model bringing great
challenges to the task of information extraction: tweets can convey unrestricted
subjects, provide little context and are written in an often informal style. No
measures have been ocially provided or adopted for a formalization of writing,
except for hashtags, that can be as irregular as the text message itself.
Thus, a complementarity can be observed between a Wikipedia that lacks real-time
updates and a Twitter that provides only unstructured data.
Twitter and Wikipedia are obviously built on fundamentally dierent principles.
Those principles are in fact what makes the two platforms so pertinent for their
usage : Twitter is extremely easy to use, everything being done to be able to
tweet from everywhere with one or two steps, allowing for astonishing network
1https://en.wikipedia.org/wiki/Wikipedia
3

4 Chapter 1 Background
dynamics. On the contrary Wikipedia edition and reading asks for more time and
more focus, leading to a highly trustable source of information.
Objections to our work will legitimately evoke the danger of denaturing the two
services by altering and mixing up these foundations, which are reasons for their
popularity and recognition. A tweet classication feature would impact Twitter's
wish of maximal simplicity; providing real-time updates to Wikipedia would be a
door open to content of a weaker quality. Exigences of neutrality must be reected
upon, statements must be correctly referenced with trusted sources, writing style
must be elaborate.
Providing a more structured model for every tweet is up to the company itself, and
would deeply question the principles that made it so popular. We believe however
that bringing real time information to Wikipedia is feasible and desirable, with a
particular attention dedicated to ensuring that its spirit is kept. Our work does
not automatically modify the articles sources, but rather lays a well identied
real-time layer on top of them, carefully considers issues of impartiality and puts
the last decision for information storage in the hands of the community.
Still growing at more than 8500 articles per day in 2012, it is however less known
that the number of active editors on Wikipedia (80 500) represent less than 0,02%
of its total monthly unique visitor count (450 million) and that this number of
new editors is declining.
2
These statistics may be interpreted in dierent ways. A
positive one might be that the growth of Wikipedia has reached its limit, for the
reason that it already describes a near satisfying amount of knowledge about our
world in its current state. Another would be that resource constraints are starting
to hinder it.
3
These interpretations however do not imply that there is no way
to alter this trend. Specically, we believe that novel features and design could
automatically enrich or assist edition, leading to a wider audience of users and
editors, drawing on the activity of Twitter's claimed 500 million users through
their 400 million tweets per day in 2012, beneting from the spread of mobile
devices.
4
Interaction between those two gigantic information worlds remains feeble. Though
Twitter is a subject appreciated by researchers, few of the publications aim at
2Statistics for Wikimedia, July 2012; an active editor is a registered user with more than 5
edits per month; http://stats.wikimedia.org/
3http://www.guardian.co.uk/technology/2009/aug/12/wikipedia-deletionist-inclusionist
4active users have been reported however to represent less
than one third of that number http://techcrunch.com/2012/07/31/
twitter-may-have-500m-users-but-only-170m-are-active-75-on-twitters-own-clients/,
http://news.cnet.com/8301-1023_3-57448388-93/twitter-hits-400-million-tweets-per-day-mostly-mobile/

Chapter 1 Background 5
constructing web applications that would enhance both, and most of the end-user
attempts did it one way. For example, Xu and Oard (2011) aims to enhance
Twitter search through a topical clustering that uses Wikipedia's information.
To the best of our knowledge, our work is unique as an initiative to conciliate
the advantages of both, and more particularly by its focus on the retrieval of
relevant tweets for Wikipedia, rather than the contrary. However, the rst part of
this project consists more of a careful review of existent technologies and services
rather than a fundamentally new algorithmic model or framework, and as such,
special attention has been given to the integration part, where general guidelines
are set, elaborate experiments described and demonstrated for some with a basic
implementation.
1.2 Related Work
1.2.1 Tweet processing
We need to link tweets to Wikipedia articles, and derive meaning from them for
further applications.
1.2.1.1 Entity Linking
In order to attain the objective of this project, a necessary step is to nd a way to
automatically establish links between tweets and Wikipedia articles. These links
are semantic, they should make a relation between the concepts that a user may
want to refer to in her 140 character microblog post and the concept material-
ized by a Wikipedia article. A natural way to do that is through the approach
called annotation: enriching text with links to topics it refers to. Annotation is
covered extensively by academic research, as we will see below. Some of those
papers particularly focus on the annotation of short text fragments, which further
corresponds to our ambition to treat tweets.
Annotation of text can be categorized in terms of the type of output that is
produced. Though other annotation types exist, we are primarily looking for a
step focused on entity linking, to directly link tweets to Wikipedia concepts.
The process of knowledge extraction that associates Wikipedia concepts to ngrams
of a text input, usually called wikication, is a particularly focused variation to

Named Entity Recognition (NER) of great interest to us. Several wikication
engines have been released, we will review the most relevant ones that have been
documented. Relatively recent, they exploit the size and quality of Wikipedia,
relying specically on the large link structure carefully selected and reviewed by
humans. In fact, most of the wikication systems rely on Wikipedia's internal
anchor set, an anchor being a portion of text that has been attributed a hyperlink
to another Wikipedia page.
The simplest approach to annotate text is to match fragments (word ngrams) with
concept names, here the Wikipedia article titles. This lexical matching, of course,
does not perform very well, as seen in the table 3 of Meij et al. (2012), yielding
poor recall scores because of the title-based only query as well as poor precision
due to the absence of a step to disambiguate concepts, explained below. The rst
wication system reported is Wikify Mihalcea and Csomai (2007). It relies on two
steps, detection and disambiguation. Detection computes the link probability score
of each ngram in the input text, and keeps those higher than a given threshold.
The disambiguation step, performed on each ngram for which an anchor has been
found, aims to select the most relevant concept among the list of concepts pointed
through this anchor. The approach taken by Wikify authors is to compare local
and global context (global refers to the entire text to annotate) with training data
from Wikipedia itself.
The authors of WikiMiner, Milne and Witten (2008b) have then built their system
on this rst achievement and added new scoring functions. They released the
software and a set of several Web services.
5
Following the approach of Bunescu
et al. (2008), their disambiguation step mainly relies on two scores, commonness
and relatedness, that will be employed as features for machine learning.
Commonness, also called prior probability (Pr), is dened with the following for-
mula :
Commonness(c, a) = Pr(c|a) =
|La,c|
|La|
Where c is a concept (represented by a Wikipedia article), a an anchor, La,c the
set of links with anchor a pointing to c and La the total set of links with anchor
a.
5http://wikipedia-miner.cms.waikato.ac.nz/demos/annotate/

Relatedness is a disambiguation criteria that selects the candidate concept that
resembles most the input text's unambiguous annotations (anchors that point to
one concept only). The similarity measure between two concepts relies on the
amount of incoming links that the two corresponding Wikipedia pages have in
common, as dened in one of their previous works, Milne and Witten (2008a).
Finally, the detection phase relies on the measure of link probability and machine
learning.
Both Wikify and WikiMiner, as well as the work of Chakrabarti et al. (2009), are
designed for annotating long and rich texts, such as encyclopedic articles or news
reports. The TAGME algorithm Ferragina and Scaiella (2012) aims to address the
challenge of wikication on short input texts, and with a complexity that allows
for real-time processing. It is also based on two steps, disambiguation and pruning.
Disambiguation can be performed as a global or local scale. Local disambiguation,
though quicker, does not achieve better results than the global counterpart Narr
et al. (2011). TAGME's disambiguation step diers from WikiMiner's in that it
takes into account the vote of all the other potential annotations, not only those
that are unambiguous, in order to compensate for the much limited context that
a short text holds.
An optimum is then found between the relatedness and commonness scores to
associate a concept to each anchor.
The pruning step computes two features, the link probability and the coherence,
and judges an annotation as irrelevant if their average is lower than a precise
threshold.
A last set of algorithms was proposed very recently to annotate microblog posts
Meij et al. (2012), CMNS-RF. The additional machine learning step (RF), yield-
ing signicant improvements over the TAGME algorithm (see table 8), involves
however a complex combination of features, and the rst step (Commonness),
though still evaluated as an improvement over TAGME, has been on the contrary
previously proven weaker Ferragina and Scaiella (2012).
As described in a benchmark by the authors of Meij et al. (2012)
6
, there exist
several other wikication algorithms, among which are Yahoo! content analysis,
Wikimeta or DBpedia Spotlight. This comparison conrms our conviction that
6edgar.meij.pro/comparison-semantic-labeling-algorithms-twitter-data/

TAGME is adapted to our project. Hence its set of algorithms has been retained
as the base for our work, as described in the following chapter link.
It can be noted that the Google Web anchor dictionary released and documented
in Spitkovsky and Chang (2012) gives hope of an extension of the anchor database
that powers all wikication algorithms cited above. Although recall would cer-
tainly benet from this, without a severe ltering step precision would be dramat-
ically reduced, as the database contains all the hazardous anchor texts written on
the entire Web.
1.2.1.2 Information extraction from micro posts
Although the wikication step already creates links between Wikipedia articles
and tweets, sucient to provide a stream for some of the applications envisioned
and implemented in this project, there is much more to be extracted from the
messages using information extraction tools, that eventually enable the construc-
tion of semantic assertions. We focus particularly on open information extraction
approaches, a subject that has received much attention lately, as they are more
adapted to the scale of the Web, and its diversity of domains.
The range of Named Entity Recognition tools (NER), to which Wikication be-
longs, is interesting to us. Some of them only perform the task of attributing
predened types to entities, revealing for example that Paris is an entity of type
Location. NER systems have reached very satisfying performance, but run into
diculties when applied to short and less formal text inputs, and particularly in
the case of tweets. Recent research eorts have however focused on this task, yield-
ing signicant improvements Ritter et al. (2011), demonstrated on the concrete
case of event extraction, cited below. NER systems can be used as a complement
to wikication, which can cast a shadow on important entities not (yet) included
in Wikipedia but that should be considered as subjects in assertions, as explained
by Lin et al. (2012). Nakashole and Weikum (2012) advocate this Dynamic Entity
Recognition.
Word sense disambiguation (WSD) algorithms precisely perform the task of entity
linking. Rusu et al. (2011) attempted to adapt two WSD algorithms to annotate
text with entities from Linked Open Data sets, resolving this issue of missing or
inadequate Wikipedia entries.

Another set of tools are dedicated to relation extraction, extracting the link be-
tween entities, as a rst step to obtain semantic predicates. The ReVerb algorithm
Fader et al. (2011), one of several works by the University of Washington's AI de-
partment
7
, relies on the OpenNLP library
8
and improves precision and recall as
compared to the standard extractors, TextRunner by Etzioni et al. (2008) and
WOE by Wu and Weld (2010). Subsequent work has again signicantly increased
performance over ReVerb Mausam et al. (2012), with an extraction not limited to
verbs but also nouns and adjectives, the addition of contextual information, with
its software released on the Web.
9
. Other works have focused on the extraction of
n-ary relations Akbik and Löser (2012).
Narr et al. (2011) have an interesting proposal, considering that everything con-
tained in a tweet that holds an entity is an annotation of that entity (objective,
subjective statements, sentiment. . . ). Their rst step is normalization Han and
Baldwin (2011) of the irregular words of the tweet.
Research has also been conducted by Rizzo et al. (2012) on the unication of NER
tools (NERD) and the translation of the results to RDF (NIF). Ngomo and Heino
(2011) unify the output of more diverse NLP tools using neural networks.
Nakashole and Weikum (2012) emphasize on the need for real-time extraction
algorithms to produce up to date assertions, also relying on the couple named
entity recognition - relation extraction.
Gerber and Ngomo (2012) built the BOA framework to infer triples from unstruc-
tured Web content by mapping natural language patterns to the predicates of
existing Linked Data. Aside from creating new sets of Linked Data, the system
exposes this database of correspondence between predicates and natural language
constructs.
Very recent works have tackled the global problem of inferring RDF/OWL linked
data from unstructured text: Augenstein et al. (2012) and Presutti et al. (2012)
10
.
Both works coordinate a series of NLP tools, that includes NER, CC and Boxer
that output Discourse Representation Structures, and WSD. These entities and
relations are then published in the form of semantic triples. These works are very
interesting for they detail a complete extraction and publishing pipeline.
7http://ai.cs.washington.edu/projects/open-information-extraction
8http://opennlp.apache.org/
9available at https://github.com/rbart/ollie
10This second work will be presented at EKAW 2012 and is not available at the moment

1.2.1.3 Network-wide information extraction
Along with the growth of Twitter as a popular and unique network has come
research eorts on its analysis, starting with Kwak et al. (2010). The new eld
of information extraction from Twitter as a whole is of particular interest to us, a
step above the complex retrieval of pertinent information from individual micro
inputs of text, involving classifying and ltering the dynamics of the network.
Earthquake detection gives a good example of a working and very useful informa-
tion extraction process Sakaki et al. (2010). Twitter trending topics (described as
being topics that are immediately popular
11
) have been watched and compared
to traditional news headlines in other media. Kwak et al. (2010) shows that more
than half of these can be categorized as timely breaking or persistent news (p.7).
Zhao et al. (2011) perform a deeper analysis of the relation between Twitter and
traditional news media. Findings particularly relevant to our project are that
their coverage of the range of categories is similar, that there are more tweets on
personal life and pop culture than on world events, that tweets can convey further
information about brands and celebrities, and nally that retweets are a way to
nd important trendy topics.
As a particularly tinted source of information, the integration of tweets to Wikipedia
would have to be rigorously framed to respect the principles evoked earlier. On the
other side of the coin, it could certainly emerge as a valuable addition, bringing
an opinion layer with information holding a focus on personal concerns, dier-
ent perspectives through unique points of view on real-world events as described
by Diakopoulos et al. (2010), a diversied coverage brought by any user speaking
about unpopular subjects or very local events Becker et al. (2011).
Research has also been conducted on the detection of news. It includes using simi-
larity measures to group similar tweets in news stories Phuvipadawat and Murata
(2010) or discover interesting tweets and reliable tweeters by mining rst the news
on traditional websites Sankaranarayanan et al. (2009). The distinction between
real-world events and Twitter-centric events has also been performed successfully
Becker et al. (2011). Or more generally, a broad classication of tweets relying
on the study of user intentions that usually follow specic patterns Sriram et al.
(2010).
11https://support.twitter.com/articles/101125-about-trends

Real world events bring up a high level of excitement on the networkpolitical
conventions, launch of new devices. . . , sometimes raising the frequency of tweets
to astonishing topical rates (e.g 2012 American election conventions reached 53
000 tweets per minute
12
). Ritter et al. (2011) have studied the extraction, ag-
gregation and classication of these events, and released an impressive Web app
demonstrating their work with an up to date calendar of incoming events.
13
Twitter has also been praised for transporting information more quickly than
conventional media, at least for the particular case of celebrity deaths Sankara-
narayanan et al. (2009), a characteristic that could make our initiative even more
relevant in the aim of bringing live information updates to Wikipedia.
1.2.2 Semantic Microblogging
Research publications have attempted to bring semantic Web concepts to mi-
croblogging. SMOB is a platform for open, semantic and distributed microblog-
ging Passant et al. (2010), that would, with wide adoption, denitely be of great
interest as a source for our system. Other projects aim at semantifying tweets
: HyperTwitter denes a syntax to enable users to make assertions (equivalence,
specialization, known and new semantic predicates) directly in tweets, and a ser-
vice that extracts them to an RDF graph Hepp (2010); TwitLogic reviews several
of those proposed nanoformat syntaxes, introducing a new and more natural one
and focusing on RDF encoding (using known vocabularies : FOAF, SIOC. . . ) and
publishing principles Shinavier (2010); the Linked Open Social Signal initiative de-
nes content encoding as well, real-time SPARQL querying, streaming, and push
notications Mendes et al. (2010). Although these publications do not help us for
the information extraction step given the low adoption of the syntaxes dened,
the semantic frameworks they dene will help us for encoding and publishing the
processed tweets.
1.2.3 Linked Data
In 2007 an extension was developed to bring semantic authoring to MediaWiki
Krotzsch et al. (2007), but was not adopted by Wikipedia
14
. A project called
12http://blog.twitter.com/2012/09/dnc2012-night-3-obamas-speech-sets.html
13http://statuscalendar.cs.washington.edu/
14Scalability reasons were evoked, but adoption of a new and more demanding syntax by the
community could also have been a reason

Shortipedia was then launched to independently capture semantic assertions from
all over the Web Vrande£i¢ et al. (2011). This has in some ways led to Wikidata,
adopted as a new project by the Wikimedia foundation and described as a semantic
knowledge base for the world.
The last proposal of this project is to leverage both Twitter and Wikipedia to
populate this promising triple repository. DBpedia would obviously be the link
between the encyclopedia and Wikidata Bizer et al. (2009).
The aim of Han and Baldwin (2011) is similar to ours in their objective of mining
information from twitter and publishing it as a semantic database, but their ob-
jectives remain vague : use the DBpedia ontology representation and augment it
with additional required structures.
1.2.4 Web page enrichment
One of the aims of this project is to automatically annotate Wikipedia pages
via tweets and the information extracted from them. This topic of Web page
enrichment regularly surfaces in research projects and commercial applications. In
fact, it has a long story, that includes the research on link injection in the context
of applying the concepts of open hypermedia to the Web Carr et al. (1998) Carr
et al. (2001), and as mentioned by Milne and Witten (2008b), simpler applications
of annotation algorithms, with Microsoft's Smart-Tag service to automatically
add links to text displayed by Internet Explorer, and Google's AutoLink toolbar
feature.
15
This acknowledgement of the need to annotate the Web was shared
by W3C's Annotea project Kahan and Koivunen (2001) and has been recently
implemented by new companies, creating for example the blog tool Zemanta
16
and a collaborative annotator iGlue
17
, followed as well by news corporations as
the BBC Kobilarov et al. (2009).
The approach of Web page annotation can be salutary as a highly formal semantic
structure could be felt as too restrictive for authors Millard et al. (2005), which
could explain the adoption issue of Semantic MediaWiki evoked above.
Extensions and widgets have been created for MediaWiki
18
, but they embed tweets
in very simple ways that mirror the ocial twitter add-ons (e.g. a simple user feed).
15Both initiatives were discontinued due to impartiality concerns coming from the public
16http://www.zemanta.com/
17http://www.iGlue.com/
18The software that powers Wikipedia http://www.mediawiki.org

Chapter 2
The Weeki system
In this chapter we present our work on the design and implementation of the
application.
2.1 Engine
We aim for a high recall retrieval of tweets, and the highest quality of enrichment.
2.1.1 Retrieval
Two possible approaches Our objective in this step is to obtain a real-time
list of tweets for each article browsed. Harnessing the principles of wikication,
there were two approaches that could lead to our expectations.
The rst one, that could be called catch them all, operates as follows: given a
concept (that is a Wikipedia article page), we want to retrieve real-time tweets
that implicitly or explicitly refer to it. Annotation methods would simply solve
the problem in the case that we could retrieve in real-time all the tweets published
on the network: they would be annotated, resulting in a relational table linking
tweets to identied concepts, that could easily be reversed to serve on-demand
the tweets relating to one concept. Unfortunately, retrieving all the tweets is not
something that we can aord for a couple of inescapable limitations in the context
of a summer project. Available technologies can now process such a big amount of
data: the software stack that we had chosen included a high performance HTTP
13

14 Chapter 2 The Weeki system
client by Twitter itself, Finagle, a NoSQL database and a language known for its
scalability (also used by Twitter), Scala
1
. The main limitations to this approach,
however, simply came from the Twitter streaming API
2
. First, only 1% of the ow
can be retrieved, leaving most of the information aside. Then came the idea to
retrieve more tweets by splitting the retrieval with the lter streaming endpoint,
but it turns out that the API allows only one simultaneous connection. Ocial
data retailers would have given us access, for a fee, to the entire feed, a load
that would have conducted us to switch to a distributed model such as Hadoop,
3
requiring signicant engineering eorts (to dene the jobs), setup, and computing
resources.
2.1.2 Linking tweets to a Wikipedia article
Hence the need for a shift in our approach, to take the problem in the opposite
way: the processes have to be red on a Wikipedia article request, making calls
for the tweets it needs only, without any global treatment or database at this
stage
4
. Tweets that relate to a concept have to be selectively retrieved using the
search API
5
. As a consequence, we need to grasp the principles of the wikication
algorithms and adapt them to serve our proper objective.
The rst step consists of knowing how to query the Twitter search API, a request
that would be made for every article served to users. The easiest way is to dene
the query as the article's title. However, given the varying writing styles of the
messages on the network and possible synonyms of this title, many tweets would
be missed.
6
Following Ferragina and Scaiella (2012) (paragraph 4), an anchor
dictionary is built using normalized article titles, redirect titles, removing single
numbers (e.g. years), single letters and anchors whose link frequency is too low.
This step can be described as a task of query expansion, as it adds to our query
string word synonyms and alternative grammatically correct writing forms. The
addition of lexical variations as described by Ritter et al. (2011) (paragraph 2.1)
would take care of the grammar errors or voluntary abreviations often found in
tweets.
1http://twitter.github.com/finagle/, http://redis.io/, http://www.scala-lang.
org/
2https://dev.twitter.com/docs/streaming-apis/streams/public
3http://hadoop.apache.org/
4A nal implementation with several users would of course need a cache for each article
5https://dev.twitter.com/docs/api/1.1/get/search/tweets
6It must be noted that Twitter obviously does not only look for search keywords in tweets.
Details about this search algorithm are nevertheless unknown.

Chapter 2 The Weeki system 15
Previous to the annotation step is the process of tokenization, that structures the
input text as a sequence of words and treats the particular features of twitter (e.g.
replacing user @ mentions by user names, removing the retweet RT keywords,
removing the # of hashtags).
The wikication algorithm described in depth in X is applied to all the tweets
retrieved. This additional step is intended to prune the irrelevant tweets caught
by this expanded search, by simply dropping those whose wikication output does
not include the original Wikipedia concept. Of course, applying the TAGME
algorithm described in 1.2.1.1 can be judged as cumbersome for this pruning task.
The advantage, however, it to fully annotate at the same time the tweets kept as
relevant with their concepts, moving us ahead to our semantic enrichment step,
for linking to Wikipedia concepts at the same time annotates the tweets with
DBpedia entities. The algorithm is besides designed for low computing times.
2.1.3 Stream Enrichment
Links embedded in a tweet are not to be ignored, as they are in some cases its
main value, the surrounding text being a description or opinion expressed about
them. In fact, according to Narr et al. (2011) 44% of tweets contain urls. A Web
service is used to resolve the shortened links and retrieve essential information
about them.
7
As a large proportion of them are articles (newspapers, blogs. . . ),
we are particularly interested in getting their title, abstract, tags and image, added
as a property of the annotated tweet object. The actual use for these links is to
turn them into article references as explained in 2.2.1.2, but it could be interesting
in future work to go further and perform a deeper analysis of the linked page. It
must also be noted that Twitter has recently released a new feature, expanded
tweets, that lets partner accounts include a link preview directly in the tweet.
8
This operation, similar to our link resolution, is at the moment rather elementary,
but could in the future take a very important role on the network. Twitter has
in fact by the past announced a new feature called annotations, that would let
users embed any kind of metadata in tweets, including microformats and links to
Wikipedia concepts, but has not been released yet.
9
Twitter hashtags are basically treated as simple words by the wikication step,
since users can omit to include words in a tweet when a hashtag is sucient to set
7Dibot, free for a limited number of monthly calls http://www.diffbot.com/
8http://blog.twitter.com/2012/06/experience-more-with-expanded-tweets.html
9www.readwriteweb.com/archives/what_twitter_annotations_mean.php

its context. These hashtags (initials for example) might not however be linked by
wikication algorithms if they do not exist as anchors in Wikipedia. The works of
Lösch and Müller (2011) and Laniado and Mika (2010) enable us to go further by
linking hashtags to Wikipedia concepts. This step comes however to the price of
higher computation time, and it usefulness could be questioned given the fact that
those hashtags not recognized by wikication algorithms often refer to in-network
discussions that could hardly be linked even manually to any Wikipedia concept.
The classication of tweets into dierent categories, namely news, events, opinions,
deals, and private messages, is interesting to us for two reasons. First, it enables
us to lter this last category, which does not correspond to the objectives of this
project. This helps us clear out the large portion of tweets that consists of daily
and personal chatter
10
. Secondly, this labeling of tweets will be used for our user
interface.
Finally, the identication of news stories and events using the methods described
in 1.2.1.3 enable us to make a specic usage of these tweets, that can be essential
side informations about concepts and are usually interesting to a large audience.
The algorithms described previously can however require the entire database of
tweets for a given period of time. This luxury that we could not aord (see 2.1.1)
would have to be reconsidered, or if possible, their methods adapted to treat only
the stream of tweets retrieved by our engine.
Natural language processing tools as well as frameworks combining them, detailed
in 1.2.1.2, enable us to derive semantic assertions (RDF triples) from individual
tweets. These systems would have to be tailored to the specicity of our in-
put, for instance by using a wikication adapted to tweets, TAGME, instead of
the Wikipedia Miner module in Augenstein et al. (2012). We aim at collecting
assertions that include Wikipedia entities, but could link them to other sets in
the fashion of the interlinking of DBpedia
11
. Possible overlapping steps in these
o-the-shelf algorithms would have to be examined in order to ensure optimal
computing time.
10Although the Wikication step already lters messages that do not contain a reference to a
Wikipedia concept
11http://wiki.dbpedia.org/Interlinking

2.2 Application Design
The description of the application that we make of this stream of tweets provided
by the engine is structured as a list of features, that happen on top of the Wikipedia
article pages. In order to respect a desire to read exclusively the original content of
a Wikipedia page, or to enjoy a distraction free experience, a slider button would
be provided to each user to permanently deactivate these features.
2.2.1 Page Enrichment
2.2.1.1 Smart Tweet Insertion
We designed an integration of tweets on the Wikipedia article based on rules of
concept matching. Tweets are inserted in well identied spots, areas of the page
where they are judged relevant.
The easiest spots to match are the existing internal links embedded in the article
text, that represent concepts. All the various section and subsection endings are
also candidate spots, with the matching performed on their headers. Their end
is targeted since they are usually written in a chronological order, while tweets
almost exclusively convey recent information.
When a match is detected, punctual spots (existing links) are lled with a small
icon, reacting to a user click by displaying the tweet in a small pop-up. Area
spots such as sections are directly followed by a related tweets link that would,
on click, unfold a horizontal list of tweets. These insertions must remain discrete
enough not to perturb the reader that just wants to focus on the encyclopedic
content of Wikipedia.
Figure 2.1: A punctual spot lled with a tweet marker (We apologize for
altering the Twitter logo, but it helped make the markers more obvious
for this experiment) (implemented feature)
As explained previously, tweets would come with a label (News, Event, Opinion)
that would dene the background color of messages as displayed on the page,
allowing users to know what kind of information they should expect from them.

Figure 2.2: The corresponding tweet rolled out on click (implemented
feature)
Figure 2.3: Area insertion (mock-up)
Moreover, our project further plans on an interaction with users reading the page,
allowing them to act on these inserted tweets. Actions would be mapped to evoca-
tive buttons such as delete, which lets users judge the insertion as irrelevant, move
to a more tted spot, and nally store, for tweets that are meaningful enough to
be permanently included in the article's tweet database or directly as Wikipedia
references (e.g. a tweet that holds value in itself as an original declaration, as the
famous https://twitter.com/BarackObama/status/241392153148915712).

2.2.1.2 Reference suggestion
Another implemented feature is an automatic and real-time suggestion of refer-
ences to users that are updating the article. Their text input is processed with
our set of algorithms in the same manner as tweets, leading to eventual matches
between extracted concepts. Links extracted from tweets, based on rules that de-
ne them as acceptable, are suggested below the input form along with an add
reference button. This feature, if carefully implemented, could help users back
up their claims more easily, taking advantage for example of the innumerable
trustworthy news article links published on Twitter.
The wealth of information extracted through natural language processing tools
(see 1.2.1.2) could in the same manner serve for a match between tweets and
article claims that have been marked as needing a citation, providing for example
a drop-down list of links identied as potential references.
12
2.2.1.3 Headline Suggestion
Research has shown how meaningful information can be extracted from Twitter
as a dynamic network 1.2.1.3. Particularly, we aim to use the extracted breaking
news stories and identied events that revolve around the concept through the
display of an informative live banner right after the article's title. To prevent the
possibility of false positives as well as correct but secondary items to be displayed,
we again rely on the individual judgments of the large Wikipedia community as
safeguards to act on these automatic headlines.
When reading an article on Wikipedia, it usually takes some time before nding
the paragraph where the most recent information is written. This headline sugges-
tion feature would enable acknowledgement of important news about this resource
at rst sight. An illustrative use case would be about athletes, whose latest com-
petition performances reported in the news are of capital interest, especially at
the time of competitions, but may not be instantly updated in the article text.
2.2.2 Linked Databases Population
The last feature dened in this project aims to employ the Wikipedia community
to judge assertions sourced from the twitter stream. Users would be proposed
12http://en.wikipedia.org/wiki/Wikipedia:Citation_needed#Citation_needed

unobtrusive pop-up asking them to conrm whether assertions appear to them
as both correct and worth of attention. These assertions could comprise multiple
disambiguation candidates that users could choose from if needed.
This feature could eventually be used as a way to populate the Wikidata project.
In order to guarantee precision, a lter could be designed to consider only asser-
tions whose predicates include properties listed in the DBpedia ontology
13
. Other
restrictions could be dened in order to respect the future Wikidata policies, for
example people involved in this process could be restricted to registered Wikipedia
users.
14
The practical feasibility of this feature would obviously depend on the performance
of the tools used upstream to infer semantic assertions, but the main bet of our im-
plementation nonetheless resides in this semi-automated process involving human
judgement as an ultimate step.
Another interesting application of this feature would be to release the set of map-
pings that have been found between natural language relations and ontology pred-
icates, similar to the BOA base described by Gerber and Ngomo (2012). The
high condence of these mappings that were validated by users would give way
to multiple applications in the domain of information extraction from the Twitter
network as well as others websites sharing this informal text style.
2.3 Implementation
Given that this project is largely based on the design of a novel application, an
early implementation was welcome as a proof of concept. The infrastructure of
Weeki, as well as two of the Wikipedia integration features described above have
been implemented: smart tweet insertion and reference suggestion. However, the
engine consists only of the wikication step. This implementation is available on
a Git repository
15
.
13http://dbpedia.org/Ontology
14http://meta.wikimedia.org/wiki/Wikidata/Preventing_unwanted_edits
15https://github.com/laem

2.3.1 Software Stack
Following the novel nature of the application, modern technologies have been
employed for its implementation.
• The programming language chosen is Scala. At the same time improving on
and compatible with Java,
16
it is an adequate choice for scalability and exten-
sibility. Brand new Web frameworks have been developed on top of it, and
it is denitely adapted to complex Web applications.
17
Its functional style
also allows for more elegant and readable programming. Last but not least,
it has been chosen by Twitter for a number of critical back-end processes
18
.
• The application is built using the Play! 2.0 framework. Now natively in
Scala, it facilitates asynchronous web programming. Play was documented
and usable enough (easy setup, well known MVC model, automatic code
reloading) for a one month deployment.
• Websocket is a new technology providing a communication channel between
browsers and servers, compatible with modern browsers.
19
It advantageously
replaces the AJAX/Comet duo in our case, and is straightforward to use with
the Play framework.
• The client is coded in the incontrovertible JavaScript and jQuery.
20
2.3.2 Extensibility
Although this implementation is only a proof of concept previewing our ideas, it
is highly extensible for future development.
Classes can simply be added to the back-end code to, for example, implement the
text to RDF feature. They could be algorithms coded in Scala or Java, or simply
Web service clients coded with the Play! helpers. The new data produced by these
additions can be passed along one of the existing Websocket channels, or simply
added to a new one. The client code is written in JavaScript, allowing the addition
16http://www.infoq.com/news/2009/07/scala-replace-java
17Notably with the Akka library http://akka.io/
18http://www.readwriteweb.com/hack/2011/07/twitter-java-scala.php
19https://en.wikipedia.org/wiki/Websockets
20This implies that clients with JavaScript disabled would not be able to use our application.
The proposed redesign of Wikipedia (evoked in future works) also currently falls short on this
constraint.

of jQuery libraries to extend functionality and interface, and enabling a future
release as a browser extension or jQuery plug-in. In order to be fully integrated to
Wikipedia, the code would have to be directly included in the MediaWiki project,
or in a more modest way as an extension.
2.3.3 Process description
The code logic that constitutes our proof of concept works as follows, triggered on
the request of a Wikipedia article. We note that the annotation Web service used
is Wikipedia Miner
21
as access to the TAGME service
22
could not be obtained.
Asynchronous calls are performed where desirable.
Figure 2.4: A screenshot of the current implementation that has just
loaded and displayed tweets
• The Wikipedia HTML page is retrieved and modied to include our CSS
les and client-side JavaScript code.
• A Twitter search API client periodically makes a search request correspond-
ing to the requested article (no query extension is performed).
• Tweets are annotated with service calls, and their URLs resolved with the
Dibot service.
23
• These enriched tweets are sent to the client through a Websocket channel,
as JSON arrays.
21http://wikipedia-miner.cms.waikato.ac.nz:8080/services/?wikify
22http://tagme.di.unipi.it/
23http://www.diffbot.com/

• The smart tweet insertion feature is performed by matching concepts and
appending HTML via jQuery code.
• The reference suggestion demo calls the server's annotation service through
a second Websocket channel and displays the corresponding list of extracted
URLs.
2.3.3.1 Anchor Dictionary
The anchor dictionary provided by the Wikipedia Miner project suces for the
query expansion task as well as the needs of the wikication algorithm.
24
However,
the Wikipedia dump from which this information is extracted dates back to July
2011, which matters given the very recent information conveyed by tweets.
We have thus attempted to build the anchor dictionary ourselves. Wikipedia
releases periodic XML dumps of its articles. The task would be to parse each
article for internal links, and store the anchor as well as the target article id. We
have used the Wikipedia Preprocessor
25
, that uses the Perl MediaWiki::DumpFile
script, to get an output le listing these anchors. A Scala script lets us create one
le for each article with its id as a name, that contains the list of anchors and
corresponding count.
This output format makes the query expansion task trivial. However, we ran
into two issues. First, as each le takes at least 4.0 ko of space on modern le
systems, the total exceeded by far the disk space given for this project. For
performance reasons, the wikication algorithm chosen would eventually need to
be implemented, and would require a list of article ids for each anchor found in
the input text. This storage method does not allow for an ecient lookup of ids.
Another storage method would hence need to be used to construct the dictionary.
NoSQL databases seem to be a good choice given the scale of the task.
24http://wikipedia-miner.cms.waikato.ac.nz/services/exploreArticle?id=
333355labels=true and http://wikipedia-miner.cms.waikato.ac.nz/services/search?
query=kiwi
25http://www.cs.technion.ac.il/~gabr/resources/code/wikiprep/

2.4 Project Management
As part of the MSc Web Technology, the successive steps of this project were
initially planned according to the following Gantt chart.
This planning was on overall respected, except for a more important overlap be-
tween research and implementation, with comings and goings between specication
and implementation.

Chapter 3
Evaluation and future work
3.1 Critical Evaluation
3.1.1 Testing
It is clear that this project is not elaborated enough to be the subject of rigorous
technical evaluation. It can nevertheless be planned to serve as guidelines for
future work.
The engine of our application that retrieves, annotates and wraps the tweets as
semantic objects would have to be tested with Information Retrieval methods.
1
Each component would have to tested against manual annotation of the tweets in
input, to be able to compute recall, precision and F1 score. While some of these
algorithms have been tested in the corresponding research paper on the case of
tweets (e.g. TAGME), a good amount of time would have to be spent on the others
(e.g. relation extraction). It is probable that recall may be aected by the fact
that the concepts of some tweets are not mentioned and hard to infer. However,
we can bet that the redundancy of messages on the network would reduce that
loss.
The second part of this work, the application idea, could not be evaluated tech-
nically as the questions to answer, such as how useful it would be ?, are more
abstract. We propose an evaluation based on a survey, that would have to be
conducted when more features would have been implemented, or at least carefully
illustrated to avoid blind answers reecting an absence of understanding from
1http://en.wikipedia.org/wiki/Precision_and_recall
25

26 Chapter 3 Evaluation and future work
a mere description of the functions. A couple of interesting questions are listed
hereafter.
• Are you an active Wikipedia contributor (more than 5 edits per month) ?
• Are you an active Twitter user ? If not, do you follow users and read their
tweets ?
• How useful would you judge the smart tweet insertion feature ? The reference
suggestion ? The headline suggestion ?
• Would you answer the factual questions (in order to enrich the Wikipedia
infoboxes / Wikidata) ?
• How do you judge the global interest of the Weeki application ?
A nal testing step would be user experience testing, to evaluate the technical
and design aspect of the implementation. Ideally, it would consist of another user
survey. Otherwise, a heuristic evaluation adapted to modern Web applications
should be applied Thompson and Kemp (2009).
3.1.2 Value and limits of our work
New natural language processing algorithms (NLP) were obviously not devel-
oped in this project, nor a framework to unify them for the purpose of our work.
NLP applied to tweets is a complex subject of study and a serious contribution to
it was not in the scope of this project. However, we did provide a corresponding
literature review that we do hope is exhaustive enough.
The principal value of this work resides in the novel idea introducedbringing
real-time updates on top of Wikipedia through the extraction of meaningful infor-
mation from Twitter. We are not aware of similar initiatives that compare to this
aspect, excepted those focusing on some parts of it and detailed in the background
chapter.
The accompanying implementation has to be seen as a proof of concept to
illustrate this idea and elementarily show its feasibility. Though it obviously lacks
nish for instance in terms of user experience and design (the reference suggestion

Chapter 3 Evaluation and future work 27
feature is hardly styled), it however provides a structure ready for future work
with a convenient and modern technology stack, and proves to be working, fast,
easy to deploy and not requiring any setup (excepted for the download of the Play
framework, the instructions are given on the Github repository).
3.1.3 Reception
Although some users of Wikipedia could judge the addition of real-time updates as
undesirable or distracting, we do believe that it would be welcome by a signicant
proportion of them, particularly by the increasing proportion of active Twitter
users that recognize and daily exploit the potential of this network.
This work primarily aims to bring attention on a new information source for
Wikipedia as well as ideas to integrate it. Depending on its reception by the
Wikipedia community, it could inspire new MediaWiki features (that wouldn't be
imposed to users, but enabled via a switch button), or interesting new browser
extensions.
3.2 Future Works
The topic range embraced by this project is large, and several aspects could not
be treated in time. Most of them are listed there.
Online Newspapers An interesting area of focus would be online newspapers.
They can be used to facilitate the detection of news that matter to Twitter users.
The Guardian and the New York Times, among others, have released semantic
Web APIs, that could be used in the entity linking step as another source of
Linked Data. The general advantage is to incorporate data sets and ontologies
that are among the most up to date, to help bridge the gap between real-time but
unstructured streams and structured but more static repositories.
2
Content specialization In the tweet query step we have focused exclusively on
content. There are however other search parameters relevant to our application:
a Wikipedia article about a music festival could in addition trigger the retrieval
2http://www.guardian.co.uk/open-platform/blog/linked-data-open-platform?CMP=
twt_gu, http://data.nytimes.com/

28 Chapter 3 Evaluation and future work
of tweets posted with a geolocation value comprised in an adequate radius, and
display a stream of tweets labeled as local. The assignation of higher weight to
tweets posted by specic users (depending on the article) could also be studied.
Personalized streams User recommendation is a dimension that was not stud-
ied in this project. Users that wish to log in with their Twitter account would see
the selection of tweets tailored to the characteristics of their account (e.g. accounts
they follow). Particular attention would have to be given to the issue of the lter
bubble eect
3
.
Wikipedia Redesign A couple of initiatives have advocated the need for a
Wikipedia interface redesign, such as the Athena proposal and the Wikipedia
Redened experiment.
4
These initiatives would bring Wikipedia closer to the
graphical and user experience level of modern websites, and we believe that our
application would be better integrated and welcome therein.
Additional sources We have focused exclusively on Twitter as an information
source, as its success makes it the most useful one, but other microblogging net-
works should eventually be taken into account, e.g. StatusNet, identi.ca and the
new App.net.
5
This could give more diversity, and maybe quality (registered users
may post more serious messages than on Twitter).
Generic applicability This project focuses on Wikipedia as a support for the
application. The idea of a retrieval, enrichment and integration of tweets relating
to a concept, however, could be of great interest for a variety of websites (e.g.
Wordpress). It could be interesting indeed to design the framework as generic,
with the integration features detailed in this paper incorporated as a component
focused on Wikipedia. An API or the ability to create other integration features
as extensions to the framework could be imagined.
3http://en.wikipedia.org/wiki/Filter_bubble
4http://www.mediawiki.org/wiki/Athena, http://www.wikipediaredefined.com/
5http://www.status.net/, https://identi.ca/ and http://www.app.net/

Chapter 4
Conclusions
We have in this paper undertaken the task of enriching Wikipedia with meaningful
information extracted from the most famous microblogging network.
A review of the state of the art information extraction tools, particularly those
more able to deal with the specicities of Twitter messages laid the foundations for
a message retrieval and enrichment engine. A particular focus was then set on the
concrete design of our application, describing four integration features enabling a
richer and live experience for reading and editing Wikipedia articles.
Moreover, an implementation was written to prove the feasibility and the interest
of our idea.
We nished with testing suggestions, a critical evaluation of our idea and work,
as well as the presentation of several diverse and challenging directions for future
work on this project.
This project has been an opportunity to explore the range of research on natural
language processing, gain theoretical and practical experience in the design of a
rich Web application, as well as for a constant and creative reection on how to
suitably propose novel features for a well established website.
29

References
Akbik, A. and Löser, A. (2012). KrakeN: N-ary Facts in Open Information Extrac-
tion. In Proceedings of the Joint Workshop on Automatic Knowledge Base Con-
struction and Web-scale Knowledge Extraction (AKBC-WEKEX), pages 5256,
Montr{é}al, Canada. Association for Computational Linguistics.
Augenstein, I., Padó, S., and Rudolph, S. (2012). LODier: Generating Linked
Data from Unstructured Text. In Simperl, E., Cimiano, P., Polleres, A., Corcho,
O., and Presutti, V., editors, ESWC, volume 7295 of Lecture Notes in Computer
Science, pages 210224. Springer.
Becker, H., Naaman, M., and Gravano, L. (2011). Beyond trending topics: Real-
world event identication on Twitter. In Fifth International AAAI Conference
on Weblogs and Social Media.
Bizer, C., Lehmann, J., Kobilarov, G., Auer, S., Becker, C., Cyganiak, R., and
Hellmann, S. (2009). DBpedia - A crystallization point for the Web of Data. Web
Semantics: Science, Services and Agents on the World Wide Web, 7(3):154165.
Bunescu, R., Gabrilovich, E., and Mihalcea, R., editors (2008). Topic Indexing
with {Wikipedia}, Menlo Park, CA, USA. AAAI.
Carr, L., Hall, W., Bechhofer, S., and Goble, C. (2001). Conceptual Linking:
Ontology-based Open Hypermedia. Proceedings of the 10th international con-
ference on World Wide Web, pages 334342.
Carr, L. A., DeRoure, D. C., Davis, H. C., and Hall, W. (1998). Implementing
an Open Link Service for the World Wide Web. World Wide Web Internet And
Web Information Systems, 1(2):6171.
Chakrabarti, S., Kulkarni, S., Singh, A., and Ramakrishnan, G. (2009). Collective
annotation of Wikipedia entities in web text. In Proceedings of the 15th ACM
SIGKDD international conference on Knowledge discovery and data mining -
KDD '09, page 457, New York, New York, USA. ACM Press.
31

32 REFERENCES
Diakopoulos, N., Naaman, M., and Kivran-Swaine, F. (2010). Diamonds in the
rough: Social media visual analytics for journalistic inquiry. In 2010 IEEE
Symposium on Visual Analytics Science and Technology, pages 115122. IEEE.
Etzioni, O., Banko, M., Soderland, S., and Weld, D. S. (2008). Open information
extraction from the web. Communications of the ACM, 51(12):68.
Fader, A., Soderland, S., and Etzioni, O. (2011). Identifying relations for open
information extraction. In Proceedings of the Conference on Empirical Methods
in Natural Language Processing, EMNLP '11, pages 15351545, Stroudsburg,
PA, USA. Association for Computational Linguistics.
Ferragina, P. and Scaiella, U. (2012). Fast and Accurate Annotation of Short Texts
with Wikipedia Pages. IEEE Software, 29(1):7075.
Gerber, D. and Ngomo, A.-c. N. (2012). Extracting Multilingual Natural-Language
Patterns for RDF Predicates. In EKAW 2012.
Han, B. and Baldwin, T. (2011). Lexical Normalisation of Short Text Messages :
Makn Sens a # twitter. Computational Linguistics, V(212):368378.
Hepp, M. (2010). HyperTwitter: Collaborative Knowledge Engineering via Twitter
Messages. In Cimiano, P. and Pinto, H., editors, Knowledge Engineering and
Management by the Masses, volume 6317 of Lecture Notes in Computer Science,
pages 451461. Springer Berlin / Heidelberg.
Kahan, J. and Koivunen, M.-R. (2001). Annotea: an open RDF infrastructure for
shared Web annotations . In Proceedings of the tenth international conference
on World Wide Web - WWW '01, pages 623632, New York, New York, USA.
ACM Press.
Kobilarov, G., Scott, T., Raimond, Y., Oliver, S., Sizemore, C., Smethurst, M.,
Bizer, C., and Lee, R. (2009). Media Meets Semantic Web How the BBC
Uses DBpedia and Linked Data to Make Connections. In Proceedings of the 6th
European Semantic Web Conference on The Semantic Web: Research and Ap-
plications, ESWC 2009 Heraklion, pages 723737, Berlin, Heidelberg. Springer-
Verlag.
Krotzsch, M., Vrandecic, D., Volkel, M., Haller, H., and Studer, R. (2007). Se-
mantic Wikipedia. Web Semantics Science Services and Agents on the World
Wide Web, 5(4):251261.

REFERENCES 33
Kwak, H., Lee, C., Park, H., and Moon, S. (2010). What is Twitter, a social
network or a news media? In Proceedings of the 19th international conference
on World wide web - WWW '10, page 591, New York, New York, USA. ACM
Press.
Laniado, D. and Mika, P. (2010). Making Sense of Twitter. In Patel-Schneider,
P., Pan, Y., Hitzler, P., Mika, P., Zhang, L., Pan, J., Horrocks, I., and Glimm,
B., editors, The Semantic Web ISWC 2010, volume 6496 of Lecture Notes in
Computer Science, pages 470485. Springer Berlin / Heidelberg.
Lin, T., Mausam, and Etzioni, O. (2012). No Noun Phrase Left Behind: Detecting
and Typing Unlinkable Entities. In Proceedings of the 2012 Joint Conference on
Empirical Methods in Natural Language Processing and Computational Natural
Language Learning, pages 893903, Jeju Island, Korea. Association for Compu-
tational Linguistics.
Lösch, U. and Müller, D. (2011). Mapping Microblog Posts to Encyclopedia Arti-
cles. Corpus, 2013.
Mausam, Schmitz, M., Soderland, S., Bart, R., and Etzioni, O. (2012). Open Lan-
guage Learning for Information Extraction. In Proceedings of the 2012 Joint
Conference on Empirical Methods in Natural Language Processing and Compu-
tational Natural Language Learning, pages 523534, Jeju Island, Korea. Associ-
ation for Computational Linguistics.
Meij, E., Weerkamp, W., and de Rijke, M. (2012). Adding semantics to microblog
posts. In Proceedings of the fth ACM international conference on Web search
and data mining - WSDM '12, page 563, New York, New York, USA. ACM
Press.
Mendes, P. N., Passant, A., Kapanipathi, P., and Sheth, A. P. (2010). Linked
Open Social Signals. In 2010 IEEE/WIC/ACM International Conference on
Web Intelligence and Intelligent Agent Technology, pages 224231. IEEE.
Mihalcea, R. and Csomai, A. (2007). Wikify! Linking Documents to Encyclopedic
Knowledge. In Proceedings of the sixteenth ACM conference on Conference on
information and knowledge management - CIKM '07, page 233, New York, New
York, USA. ACM Press.
Millard, D. E., Gibbins, N. M., Michaelides, D. T., and Weal, M. J. (2005). Mind
the semantic gap. In Proceedings of the sixteenth ACM conference on Hypertext

34 REFERENCES
and hypermedia - HYPERTEXT '05, page 54, New York, New York, USA. ACM
Press.
Milne, D. and Witten, I. H. (2008a). An eective, low-cost measure of semantic
relatedness obtained from Wikipedia links. In Proceedings of the rst AAAI
Workshop on Wikipedia and Articial Intelligence.
Milne, D. and Witten, I. H. (2008b). Learning to link with wikipedia. In Proceeding
of the 17th ACM conference on Information and knowledge mining - CIKM '08,
page 509, New York, New York, USA. ACM Press.
Nakashole, N. and Weikum, G. (2012). Real-time Population of Knowledge Bases:
Opportunities and Challenges. In Proceedings of the Joint Workshop on Au-
tomatic Knowledge Base Construction and Web-scale Knowledge Extraction
(AKBC-WEKEX), pages 4145, Montr{é}al, Canada. Association for Compu-
tational Linguistics.
Narr, S., De Luca, E. W., and Albayrak, S. (2011). Extracting semantic an-
notations from twitter. In Proceedings of the fourth workshop on Exploiting
semantic annotations in information retrieval - ESAIR '11, page 15, New York,
New York, USA. ACM Press.
Ngomo, A.-c. N. and Heino, N. (2011). Federated Knowledge Extraction for Se-
mantic Web Applications. ACL.
Passant, A., Bojars, U., Breslin, J. G., Hastrup, T., Stankovic, M., and Laublet,
P. (2010). An Overview of SMOB 2: Open, Semantic and Distributed Mi-
croblogging. In Cohen, W. W. and Gosling, S., editors, ICWSM. The AAAI
Press.
Phuvipadawat, S. and Murata, T. (2010). Breaking News Detection and Track-
ing in Twitter. In 2010 IEEE/WIC/ACM International Conference on Web
Intelligence and Intelligent Agent Technology, pages 120123. IEEE.
Presutti, V., Draicchio, F., and Gangemi, A. (2012). Knowledge extraction based
on Discourse Representation Theory and linguistic frames. In EKAW 2012.
Ritter, A., Clark, S., and Etzioni, O. (2011). Named Entity Recognition in Tweets:
An Experimental Study.
Rizzo, G., Troncy, R., Hellmann, S., and Bruemmer, M. (2012). {NERD}
meets {NIF}: {L}ifting {NLP} extraction results to the linked data cloud. In

REFERENCES 35
{LDOW}, 5th {W}orkshop on {L}inked {D}ata on the {W}eb, {A}pril 16,
2012, {L}yon, {F}rance, {L}yon, {FRANCE}.
Rusu, D., Fortuna, B., and Mladenic, D. (2011). Automatically Annotating Text
with Linked Open Data. In Bizer, C., Heath, T., Berners-Lee, T., and Hausen-
blas, M., editors, 4th Linked Data on the Web Workshop (LDOW 2011), 20th
World Wide Web Conference (WWW 2011)., Hyderabad, India.
Sakaki, T., Okazaki, M., and Matsuo, Y. (2010). Earthquake shakes Twitter
users: real-time event detection by social sensors. In Proceedings of the 19th
international conference on World wide web - WWW '10, page 851, New York,
New York, USA. ACM Press.
Sankaranarayanan, J., Samet, H., Teitler, B. E., Lieberman, M. D., and Sperling,
J. (2009). TwitterStand. In Proceedings of the 17th ACM SIGSPATIAL Inter-
national Conference on Advances in Geographic Information Systems - GIS '09,
page 42, New York, New York, USA. ACM Press.
Shinavier, J. (2010). Real-time #SemanticWeb in = 140 chars. IN PROCEED-
INGS OF LINKED DATA ON THE WEB 2010, ON WWW2010.
Spitkovsky, V. I. and Chang, A. X. (2012). A Cross-Lingual Dictionary for English
Wikipedia Concepts. In LREC 2012 Istanbul.
Sriram, B., Fuhry, D., Demir, E., Ferhatosmanoglu, H., and Demirbas, M. (2010).
Short text classication in twitter to improve information ltering. In Proceeding
of the 33rd international ACM SIGIR conference on Research and development
in information retrieval - SIGIR '10, page 841, New York, New York, USA.
ACM Press.
Thompson, A.-J. and Kemp, E. A. (2009). Web 2.0: extending the framework
for heuristic evaluation. In Proceedings of the 10th International Conference NZ
Chapter of the ACM's Special Interest Group on Human-Computer Interaction
- CHINZ '09, pages 2936, New York, New York, USA. ACM Press.
Vrande£i¢, D., Ratnakar, V., Krötzsch, M., and Gil, Y. (2011). Shortipedia aggre-
gating and curating Semantic Web data. Web Semantics: Science, Services and
Agents on the World Wide Web, 9(3):334338.
Wu, F. and Weld, D. S. (2010). Open information extraction using Wikipedia. In
48th Annual Meeting of the Association for Computational Linguistics, pages
118127. Association for Computational Linguistics.

36 REFERENCES
Xu, T. and Oard, D. W. (2011). Wikipedia-based topic clustering for microblogs.
Proceedings of the American Society for Information Science and Technology,
48(1):110.
Zhao, W., Jiang, J., Weng, J., He, J., Lim, E.-P., Yan, H., and Li, X. (2011).
Comparing Twitter and Traditional Media Using Topic Models. In Clough, P.,
Foley, C., Gurrin, C., Jones, G., Kraaij, W., Lee, H., and Mudoch, V., editors,
Advances in Information Retrieval, volume 6611 of Lecture Notes in Computer
Science, pages 338349. Springer Berlin / Heidelberg.

Weeki - Wikipedia <- tweets

More Related Content

Similar to Weeki - Wikipedia <- tweets (20)

Recently uploaded (20)