Annotating Real Time Twitter’s Images/Videos Basing on Tweets

Mohamed Kharrat, Anis Jedidi and Faiez Gargouri

MIRACL Laboratory, University of Sfax, Sfax, Tunisia

Keywords: Twitter, Annotation, Image, Video.

Abstract: Nowadays, online social network “Twitter” represents a huge source of unrefined information in various

formats (text, video, photo), especially during events and abnormal cases/incidents. New features for

Twitter mobile application are now available, allowing user to publish direct photos online. This paper is

focusing on photos/videos taken by user and published in real time using only mobile devices. The aim is to

find candidates for annotation from Tweet stream, then to annotate them by taking into accounts several

features based only on tweets. A preprocessing step is necessary to exclude all useless tweets, we then

process textual content of the rest. As a final step, we consider an additional characterization (spatio-

temporal and saliency) to get outcome of the annotation as RDF triples.

1 INTRODUCTION

Twitter is a microblogging service which enables

people to share between them not only short

messages but also multimedia contents called:

Tweets. Tweets can include a lot of true and/or false

information. Twitter has become an important

source for news by reporting real-world

events/incidents.

Our concern is about shared pictures/video which

represent additional information to understand what

user wants to say visually and intuitively.

Our work is motivated by the need of annotating

real time and real world image/video. Those can be

efficiently used in news and are required in

applications that cannot afford the complexity and

associated time with current image processing

techniques.

On the other hand, Twitter provides unrefined

data, in a timely manner so information is spreads

incredibly fast and is posted before it makes it into

official and suitable resources for knowledge

extraction.

In this paper we address the question whether we

can exploit or not this social media to extract new

facts/news based on shared images or videos. We

present approaches for the task of social image/video

annotation. The proposed methods are based only on

the tweets accompanying shared image, without the

use of ’slow levels features.

We primarily use tweet level features and

partially user level as authors of (Gupta, 2013)

proved the primer performs the best accuracy. Then,

only the evaluation make possible to distinguish the

best combination of all used features.

Through the process presented in the next section

we aim to a comprehension of what is happening in

one’s environment. Peoples , who find themselves in

abnormal circumstances, can describe the current

situation in real time with details using on-site

information such as what is happening, where, when

and who is involved.

Following (Feng, 2008) in this study, we propose

that an image/video can be annotated with

keywords, visual named entities and semantics

interpreted attributes. We hypothesize that tweets

containing an image/video are more likely to contain

on-the-ground information taken and shared by

eyewitnesses. Candidate tweets should be closer to

the abnormal situation. In fact on-the-ground

information tends to contain highly informative

value.

Real-time stream of information provided by

Twitter can be accessed via a single API. In addition

a rich variety of sources publish information via

Twitter like traditional media or citizen journalists

(Hermida 2010). Tweets also contain metadata that

can be exploited, like location, hashtags and user

profile information. But the biggest drawback of

Twitter is it noisy and unrefined shared data. Tweets

can be real news but also rumors or spam.

Kharrat, M., Jedidi, A. and Gargouri, F..

Annotating Real Time Twitter’s Images/Videos Basing on Tweets.

In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 1: KDIR, pages 293-300

ISBN: 978-989-758-158-8

293

Results of this study will help us to state if

shared image/video within tweets indicate on-the-

ground information and if they are suitable for use as

news?

Current approaches exploiting social media,

focus primarily on large scale incidents like

earthquakes which are characterized by a big

number of tweets as well as users interactions.

Meanwhile, small-scale incidents/events have

usually a small number of tweets and less user’s

interactions. This is challenging us for detecting

image/video candidates for annotation.

Compared with event detection in news texts,

Twitter provides more opportunities and challenges.

Authors of (Tov, 2011) reported that Twitter can

broadcast news faster than traditional media (except

those shared and published by websites).

Actually, during any small incidents/events,

thousands of microblogs (tweets) are posted in short

intervals of time. Typically, only a small fraction of

these tweets contribute to share new events and

abnormal activities, while usual users simply share

their sentiments or opinions. Real-time processing of

tweets contributing to supply news is very important

before it spreads widely via Internet.

Consequently, automatic differentiating of

relevant tweets, to be candidates for annotation from

those reflecting opinions/sentiments, is a non-trivial

challenge, mainly because of tweets characteristics

which include usually emoticons, abbreviations,

question marks... We apply in our approach, Natural

Language Processing (NLP) techniques to address

this challenge to annotate shared image/video by

user in their tweets. In addition, we apply several

steps to detect candidates for annotation, employing

filtering techniques to remove spurious

events/incidents. Finally we extract terms and facts

from those tweets which best characterize the

situation, represent visually the image/video and are

most efficacious in retrieving news.

Figure 1 shows the outlines of the baseline system

steps described below.

Figure 1: Outlines of the baseline system.

The rest of this paper is organized as following:

Section 2 presents background and definitions. In

Section 3, we discuss related work in the area of

twitter processing and detecting. Section 4 and 5

describe the pre-processing step and used features

respectively. We introduce the baseline system in

section 6. Finally, section 7 presents our conclusion.

2 BACKGROUND AND

DEFINITIONS

Twitter is a social media network, where users

follow other users in order to receive information

along timeline. Such information could be small text

messages called tweets including multimedia

resources as images or videos. There are also

relationships between users, which mean each user

has followers and followees. Tweets can be

republished any time throughout the network, this

operation is called re-tweeting. A retweeted message

usually starts with “RT @username”, where the @

sign represents a reference to the user who originally

published the message. Users could also use

hashtags (#) to identify certain topics. Hashtags are

similar to tags. The most used hashtags in the

network become trending topics reflecting incidents/

events.

3 RELATED WORK

Through the process presented below, we are able to

perform annotation of images/videos from

unrestricted domains using only content of tweets.

Most of previous work on Twitter was done on

event detection and most of them detect only

specific types of events (Qin, 2013) likewise

earthquake events detection from Twitter (Sakaki,

2010).

Authors of (Corvey, 2012) analyzed one of the

major aspects of applying computational techniques

and algorithms to social media data in order to

obtain useful information i.e. linguistic and

behavioral annotations.

Authors of (Dave, 2010) reported that

classification performance is significantly degraded

in the more practical cross domain classification.

This practical limitation motivated us to avoid using

classification, in addition to the fact that we cannot

limit the domain, since we are not detecting events,

but images/videos about small incidents, events…

Majority of the previous works on event

detection using social media has focused on using

topic detection methods to identify breaking news

stories. Streaming document similarity measures

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

294

(Petrovic, 2010), (Osborne, 2014) and online

incremental clustering (Becker, 2011) have been

shown to be effective for this purpose. Other

approaches have aimed to pick up more localized

events. These have included searching for spatial

clusters in tweets (Osborne, 2013), leveraging the

social network structure (Aggarwal, 2012),

analyzing the patterns of communication activity

(Chierichetti, 2014) and identifying significant

keywords by their spatial signature (Abdelhaq,

2013).

Work has been done to extract situational

awareness information from the vast amount of data

posted on OSM (Online Social Media) during real-

world events. (Nicholas, 2015) adapt existing bio-

surveillance algorithms to detect localized spikes in

Twitter activity which could be classified as real

events with a high level of confidence.

Regarding tweets processing, the biggest

problems are their contents and their important

amount which are considered as an important

resource that can play a critical role in crisis (Palen,

2010). However working on tweets, needs also a

filter step because many of them are fake or rumors

(Palen, 2010). In fact, authors of (Gupta, 2013)

proposed a solution to characterize and identify the

propagation of fake pictures on online social media

during Hurricane Sandy.

(Li, 2011) introduced a system for searching and

visualization the tweets related to small scale

incidents, based on keyword, spatial, and temporal

filtering.

Most existing approaches are focused on

analyzing text-based messages from Twitter.

Meanwhile multimedia (image/video) based

approaches have not been extensively addressed

hitherto despite the fact that real time images/video

shared on social media can refer to valuable

information towards improving news. This work

tries to confirm this hypothesis by exploiting this

type of media through tweets.

Indeed, we hypothesize that relevant tweets

containing an image/video are more likely to contain

on-the-ground information – for example photos

taken and posted by eyewitnesses. Consequently

such tweets should be closer to an incident/event, i.e.

affected areas. Therefore, we investigated the

following research questions: Does the existence of

image/video within tweets indicates on-the-ground

information and is thus suited for news?

Few studies interested in annotation

images/video shared on twitter or extracting

information from them. Most of them use low levels

features which could never be good for real-time

processing. Several of them on the contrary, use

external photos from other social media to annotate

tweets like the approach proposed in (McParlane,

2014) by exploiting Twitter and Wikipedia for the

annotation of event images.

In addition, there is a lot of research work

analyzing accompanied text on pictures (Srihari,

1994) but, there has been less works on detecting

events or analyzing contents of images shared on

Twitter.

Authors in (Raad, 2014) proposed an application

where events are detected from photos, capturing the

Where, When and Who dimensions of an event, and

describing (temporal, spatial, and semantic)

relationships between events using only image

metadata. Another work for Automatic image

annotation using auxiliary text information was

presented in (Feng, 2008).

In (Phuvipadawat, 2010) authors proposed an

approach to find the most frequent image related to a

tweet and return it from internet.

In (Leong, 2010) authors introduced extractive

approach for automatic image tagging by natural

language resources for processing texts surrounding

images. However this approach used Flickr

repository and Wikipedia which could not be

convenient for tweets. Finally, advanced approach in

(Chen, 2013) aims to mine salient images related to

one specific object by proposing an image clustering

and ranking algorithm.

4 PREPROCESSING

We describe here, several steps to apply on tweet

stream to eliminate all useless tweets.

To capture duplicated tweets, it is possible to

modify the pre-processing also in order to

duplication capture, e.g. one could filter out the

“RT” string and the user mentioning and repeating

the same hashing procedure; or one could detect

near duplicates using Jaccard similarity (using also

an inverse index for speed).

In fact, we do not exploit redundant/duplicated

tweets. If we consider a tweet as relevant candidate

for annotation, then if the same information is shared

and published after a while (out of window time), it

will not be considered as new anymore but old one.

Otherwise, we take in consideration redundant

tweets if only they were shared in the window

interval, which suppose that users are in the same

location and watching the same incident/event.

To avoid spam tweets from bots, and basing on

the tweeting interval (Chu, 2010), we can detect

Annotating Real Time Twitter’s Images/Videos Basing on Tweets

295

automated users who tend to have a periodic or

regular timing of tweets. Though, it is possible that

bots overcome such detection but it still an

additional filter for us.

In the same context, non-spammers users spend

more time interacting with other users (Benevenuto,

2010). In order to exclude tweets of spammers users,

we looked at a several of features of them, like the

number of times the user was mentioned by other

users and number of times the user was replied.

All tweets not containing real images or video

taken from mobiles will be eliminated.

All tweets containing URLs or shared image or

video from Internet or other mobile apps will also be

eliminated.

Further, we remove tweets automatically

generated by check-in services such as Swarm by

detecting the patterns “I’m at” and “mayor”.

In fact, Twitter applications designed specifically

for mobile devices (e.g., twitter for iPhone/Android)

are frequently used in author tweets and used by

individual persons. Organizations, unlike, primarily

use the Twitter web version and content

management software applications to publish and

manage content on Twitter (De Silva , 2014)

Terms with less than three characters: trigrams

will be eliminated. stopwords and performing POS

tagging will be removed as well as tweets containing

smileys (mdr, ptdr…).

We consider only English tweets reducing thus

the number of tweets that need to be processed in

further steps.

Last process is to convert slang words to real

words using slang dictionary, i.e convert “abwt” to

“about”.

5 FEATURES

This section describes the linguistic features used for

distinguishing candidate tweets or to filter non

eligible tweets.

The parts-of-speech (POS) tags are identified

using a probabilistic tokenizer and POS tagger

designed explicitly for tweets (Owoputi, 2013),

which can also identify emoticons and exclamations.

The required feature extraction step, is based on

the content and the user information of tweets.

Those are summarized in Table 1 and described in

the following paragraph.

Table1: Content and user features.

Content features

Length of tweet

Contains question marks

Number of retweets

Number of hashtags

Number of Favorite

User features

Number of followers

Number of tweets

Number of following friends

Follower-friend ratio

5.1 Content Features

These features are solely based on the content of

tweets. We rely on the features used by Gupta et al.

(Corvey, 2012), to which we add the number of

retweets for each tweet. The features are listed in

Table 1. Starting from the tweet characteristics, we

compute features such as the length of the tweet and

the number of words it contains. Also, we include

features such as the number of question marks (as

part of expressing shock, disgust…) and exclamation

marks and exclamatory words (e.g., :(, ‘omg!’, ‘oh

no!’). We expected incident tweets to contain

exclamations much more frequently than ordinary

tweets.

We take into account the sentiment of the tweet

which usually contains a higher fraction of

‘subjective’ words, relying on a predefined list of

strongly subjective words and subjectivity lexicons

(Volkova, 2013) specifically developed for tweets.

We compute the number of subjective words it

contains.

Unlike headline verbs which tend to be more

formal, personal verbs tend to represent personal

activities, communications, and emotions. We use a

verbs list identified by authors of (De Silva , 2014)

containing 2221 personal English verbs.

After detecting the text language of each of the

tweets using an open language detection library, a

list of tweet features for each Tweet of the stream is

produced.

5.2 User Features

We also extract features from the Twitter user, the

author of the post. It includes user’s number of

friends and followers (In fact there are opportunistic

users that follow important number of people in

order to be followed back), as well as frequency of

tweeting.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

296

6 BASELINE SYSTEM

We try to find the names of persons who are both

visible in the image/video and described in the

tweet. A baseline system will be started when all

persons found in the text are assumed present in the

shared image/video. This assumption depends of the

precision of the used NER (Named Entity

Recognition) and percentage of tweets that discuss

people not present in the image. In some cases this

problem can be solved by using a pattern of

syntactic formulas but developing a system that

could extract this information is not trivial, and even

so only a very small percentage of the texts in our

test corpus contain this kind of information.

6.1 Mechanism

Most of tweets are not real world stories, but rather

talks about personal life, conversations, expressing

humor, or spam. Running a first story detection or

event detection system on this huge data would yield

an important amount of new stories every day, most

of which would be interesting only for few people.

However, when something significant happens (e.g.,

a minister seen in a restaurant), a lot of users write

about this, either to share their opinion or just to

inform others of the situation. Our goal here is to

automatically detect these significant candidates (i.e

Figure 2) for annotation, with a minimal of

information.

Figure 2: Example of candidate Tweet.

We considered bursts of tweets what appear after

an event happens. We exploited this fact and tried to

measure the number of times the given entities refers

to the main tweet on Twitter recently.

Over 1 hour of fixed window time, and after a

tweet have been trigged the process by analyzing the

sudden peaks (number of retweet and of replies

reached), we crawled the follower list of all the

unique users that had tweeted the original tweet to

compute the overall of followers users and those

who retweet without following the original user.

We did not consider all retweets of the original

one, because according to research in (Gupta, 2013)

86% tweets spreading the fake images were

retweets, hence very few were original tweets. In

fact, at crisis time, users retweet anything from each

other regardless of the fact whether they follow them

or not.

Thus, the semantic representation of a tweet “t”

consists of keywords (KW(t)) and keyphrases

(KT(t)).

In fact segments can be advantageous for tweet

processing as they have much smaller quantity than

tweets itself and are more semantically, more

meaningful than keywords (Li, 2011).

We extracted the longest sequence of nouns as

well as proper nouns/Named entities and keyphrases

using POS tagger designed explicitly for tweets

(Owoputi, 2013).

We utilized lemmatized terms instead of raw

terms through Wordnet

We considered nouns, hashtags and proper nouns

as keywords.

Next, a graph based tweets was created

consisting of nodes which represents tweets while

the similarity between two tweets is represented by

the edge as weight. For similarity between two

tweets, t1 and t2 we adopt Eq. 1 proposed by

(Panem, 2014) authors with a small adaptation.

Sim(t1,t2)=w x sim(KT(t1), KT(t2)) + (1-w)

x sim (KW(t1), KW (t2))

(1)

Where

Sim (KT(t1), KT(t2))= |KT(t1) ᴖ KT(t2)| (2)

And

Sim (KW(t1), KW(t2))= |KW(t1) ᴖ KW(t2)| (3)

Here, “w” denotes the weight given to the

keyphrases and (1 − w) denotes the weight given to

the keywords. In our experiments, “w” will take at

the beginning many values in order to determine the

best one.

Stream of tweets is unbounded, for this reason

we do not store all the previous data in main

memory nor compare the new document to all the

previous tweets.

Automatic distinguish between a candidate tweet

or not, should be made in bounded time (preferably

constant time per document), and using bounded

Annotating Real Time Twitter’s Images/Videos Basing on Tweets

297

space (also constant per document). We have chosen

1 hour as property of our approach.

This property allows us to avoid limiting the

number of documents inside a single bucket to a

constant. If the bucket is full, the oldest document in

the bucket is removed. Note that the document is

removed only when it exceed 1 hour. Note that this

way of limiting the number of kept documents is in a

way topic-specific.

To achieve that, each tweet should be processed

as it arrives with building up dynamically tweet

clusters representing events. In particular, for each

incoming tweet, it should be compared against the

stream of previously seen tweets using a fast hashing

strategy

If the current tweet is sufficiently (textually)

dissimilar from its nearest neighbor, it is flagged.

The system attempts to reduce false positives by

waiting for a very short deferral period of 1 hour,

thus it can collect all follow-up posts and produce

clusters of closely related tweets. For an event

cluster, the tweet closest to the centroid of the

cluster (using a standard vector space) is emitted.

We boost the terms that correspond to named

entities and hashtags by some constant factor

6.2 Additional Characterization

6.2.1 Saliency

We use salience to determine what terms will be

used for annotations and included in triples.

We define here, salience measure, which

represents a value between 0 and 1 that refers the

importance of an entity in a tweet. Usually it is

calculated simply as tf (term frequency) x idf

(inverse document frequency) of terms that represent

the entity in the text.

However, for tweets which are short texts,

another measure is necessary, because almost all

entities are only mentioned once so we need a

reliable way to discern their salience.

To determine which terms are salient and

describing well the shared image/video, we calculate

the co-occurrence of terms and its variants, for

candidate tweets and its replies. We estimate the

term co-occurrence statistics with the user frequency

of a term: the number of people using that term in a

given location. In this case, the user frequency is not

significant, so term co-occurrence is computed

within the term frequency of all similar tweets.

Term co-occurrence is traditionally computed as

the number of times term ‘t’ and term ‘w’ appear in

the same tweet TW, divided by the number of times

term ‘t’ and ‘w’ appear in any tweet in the same

window time and in close location.

Some users of twitter are very active, so they

may generate a lot of tweets in a small interval of

time. Term frequency may produce an estimate of

the term distribution biased toward a particular user

or set of users. To avoid dominating messages from

one user, we estimate the term co-occurrence with

the user frequency. This is proved to be efficient by

author in (O’Hare, 2013)

For better understanding, let’s consider the

following scenario. Assume that a user took with his

smart phone a photo of Michael Schumacher leaving

the Grenoble hospital after 3 years. There are two

cases, conditional by enabling the GPS on this

phone: Or the user tags the tweet with GPS

location/coordinates, either the GPS receiver of the

phone calculates the latitude and the longitude of the

location and the data are stored in the Exif descriptor

of the shared photo/video.

6.2.2 Geolocation

Then, the system applies a geobased search on

following tweets to find user-tagged tweets taken by

other users in the surroundings of this hospital by

applying similarity between tweets.

We identify the tweets about the same topics by

looking for same terms. We compute the term co-

occurrence between terms in the main tweet, and the

terms that occur in following tweets.

From the retrieved set, the system begins

processing to select the tags of the visually matching

image/video which can be used to produce

annotation.

There is difference in case of tweets which are

not geotagged. Indeed, in such case, we identify

locations using Stanford NER. Secondly, to relate

the location mention to a point where the incident

happened, we geocode the location strings. In this

case, we create a set of word unigrams, bigrams, and

trigrams. These are sent to the geographical database

GeoNames to identify city names in each of the n-

grams and to extract geocoordinates. As city names

are ambiguous around the world, we choose the

longest n-gram as the most probable city.

6.2.3 Temporal

To identify temporal dimension of a tweet, we adopt

the HeidelTime (Jannik, 2010) framework for

temporal extraction. HeidelTime is a rule-based

approach that extracts temporal expressions from

text documents and normalizes them according to

the TIMEX3 annotation standard.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

298

6.2.4 Facts Extraction Triples

A metadata, describing multimedia resource

image/video, consists of a set of attributes. We

formally represent a metadata of this resource as:

res: (G, T, O, F) where:

• G represents the geo-location of the resource

when captured (i.e., GPS coordinates or

Geoname),

• T represents the creation date/time of the

resource,

• O represents a set of objects of interest identified

in the resource (i.e. person, monument….).

• F represents facts

Let ‘res ∈ I’ be an image or video shared by a

user. Formally, we represent a resource as follows:

res: (resid, meta)

Where:

• resId: is the identifier of the resource,

• meta: is the metadata describing the resource.

We use (REVERB), open information extraction

system (Fader, 2011), to filter out relevant facts as

RDF triples. We evaluate facts in terms of their

well-formedness, their correctness, and their

relevance.

RDF Subjects describe agents, causers, or

experiencers, while RDF objects describe entities

undergoing a state of change or being affected by an

action.

Triples are characterized by (i) the first argument

of the extracted triple is one of the named entities in

candidate tweet, (ii) the most frequent sense of the

verb has the super sense stative, possession,

competition or creation according to WordNet, and

(iii) t none of the arguments are stop words. We then

extract triples made up of a verb and the head words

of the two arguments.

After creating RDF subject and objects

corresponding to RDF triples, not of the main tweet

but of related tweets (replies), we select only

candidates named entity that have high salience.

7 CONCLUSIONS

Social media networks, particularly Twitter, become

strong and fast broadcaster of news. In this paper,

we propose an approach to annotate automatically

multimedia shared documents (videos or images) in

Twitter social media. It corresponds to real world

resources taken and shared only by mobile devices,

not using web interface nether sharing applications.

We aim to enhance news with unshared and

unpublished incidents and events. Before the

annotation we detect candidates for annotation, and

then we apply several steps processing contents of

tweets in order to produce annotation as RDF triples.

After crawling tweets using Twitter API, which

return about 1% of all current tweets, next step will

be the evaluation of this approach.

REFERENCES

Abdelhaq, H., Sengstock, C., and Gertz, M., 2013.

Eventweet: Online localized event detection from

twitter. In Proceedings of the VLDB Endowment

6(12):1326–1329.

Aggarwal, C. C., and Subbian, K., 2012. Event detection

in social streams. In SDM, volume 12, 624–635. SIAM.

Benevenuto, F., Magno, G., Rodrigues, T., Almeida, V.,

2010. Detecting Spammers on Twitter. In

Collaboration, Electronic messaging, Anti-Abuse and

Spam Conference.

Becker, H., Naaman, M., Gravano, L., 2011. Beyond

trending topics: Real-world event identification on

twitter. In Proceedings of the Fifth International

Conference on Weblogs and Social Media, Barcelona,

Catalonia, Spain. ICWSM 11:438–441.

Chen, X., Chen, M., Kim, EY., 2013. Mining salient

images from a large-scale blogosphere. In Proceeding

of 8th International Conference for Internet

Technology and Secured Transactions (ICITST).

Chu, Z., Gianvecchio, S., Wang, H., and Jajodia, S., 2010.

Who is tweeting on Twitter: human, bot, or cyborg?.

In Proceedings of the 26th Annual Computer Security

Applications Conference Austin, Texas: ACM, 2010.

Chierichetti, F., Kleinberg, J., Kumar, R., Mahdian, M.,

and Pandey, S. 2014. Event detection via

communication pattern analysis. In Eighth

International AAAI Conference on Weblogs and Social

Media.

Corvey, WJ., Verma, S., Vieweg, S., Palmer, M., and

Martin, JH., 2012. Foundations of a multilayer

annotation framework for twitter communications

during crisis events. In Proceedings of the Eight

International Conference on Language Resources and

Evaluation (LREC’12), Istanbul, Turkey.

Dave, KS., Varma, V., 2010. Pattern Based Keyword

Extraction for Contextual Advertising. In Proc. of the

19th ACM Intl. Conf. on Information and Knowledge

Management (CIKM), pages 1885–1888.

De Silva, L., Riloff, E., 2014. User Type Classification of

Tweets with Implications for Event Recognition. In

Proceedings of the Joint Workshop on Social

Dynamics and Personal Attributes in Social Media.

Fader, A., Soderland, S., Etzioni, O., 2011. Identifying

Relations for Open Information Extraction. In

Proceedings of the Conference of Empirical Methods

in Natural Language Processing, EMNLP.

Annotating Real Time Twitter’s Images/Videos Basing on Tweets

299

Feng, Y., and Lapata, M., 2008. Automatic image

annotation using auxiliary text information. In

Proceedings of the Association for Computational

Linguistics.

Hermida, A., 2010. From TV to Twitter: How Ambient

News Became Ambient Journalism. In Media/Culture

Journal, Vol. 13, No. 2, May 2010.

Jannik, S., Gertz, M., 2010. HeidelTime: High Qualitiy

Rule-based Extraction and Normalization of Temporal

Expressions. In SemEval'10.

Gupta ,A., Lamba, H., Kumaraguru, P., Joshi, A., 2013.

Faking Sandy: Characterizing and Identifying Fake

Images on Twitter during Hurricane Sandy. In

Proceedings of the 22nd international conference on

World Wide Web companion.

Li, R., Lei, K.H., Khadiwala, R., Chang, K.C.C., 2011.

Tedas: A twitter-based event detection and analysis

system. In 11th International Conference on ITS

Telecommunications (ITST).

McParlane, PJ., Jose, J., 2014. Exploiting Twitter and

Wikipedia for the Annotation of Event Images. In

Proceedings of the 37th international ACM SIGIR

conference on Research & development in information

retrieval.

Nicholas A. Thapen, Donal Stephen Simmie, Chris

Hankin. 2015. The Early Bird Catches The Term:

Combining Twitter and News Data For Event

Detection and Situational Awareness. In Journal:

CoRR Vol. abs/1504.02335.

Leong, CW., Mihalcea, R., and Hassan, S., 2010. Text

Mining for Automatic Image Tagging. In Proceedings

of the 23rd International Conference on

Computational Linguistics.

Osborne, M., Moran, S., McCreadie, R., Von Lunen, A.,

Sykora, M. D., Cano, E., Ireson, N., Macdonald, C.,

Walther, M., and Kaisser, M. 2013. Geo-spatial event

detection in the twitter stream. In Advances in

Information Retrieval. Springer. 356–367.

Ounis, I., He, Y., et al., 2014. Real-time detection,

tracking, and monitoring of automatically discovered

events in social media. In Proceeding of the 52nd

Annual Meeting of the Association for Computational

Linguistics: System Demonstrations, Baltimore, MD,

USA, 23-24 Jun 2014, pp. 37-42.

Owoputi, O., O’Connor, B., Dyer, C., Gimpel, K.,

Schneider, N., and Smith, NA., 2013. Improved part-

of-speech tagging for online conversational text with

word clusters. In Proceedings of NAACL-HLT, 2013,

pp. 380–390.

O’Hare, N., Murdock, V., 2013. Modeling locations with

social media. Journal of Information Retrieval, 16(1).

Palen, L., Anderson, KM., Mark, G., Martin, J., Sicker, D.,

Palmer, M., and Grunwald, D., 2010. A vision for

technology-mediated support for public participation

& assistance in mass emergencies & disasters. In

Proceedings of the 2010 ACM-BCS Visions of

Computer Science Conference, ACM-BCS ’10.

Panem, S., Bansal, R., Gupta, M., Varma, V., 2014. Entity

Tracking in Real-Time using Sub-Topic Detection on

Twitter. In Proceedings of 36th European Conference

on IR Research, ECIR 2014, Amsterdam.

Petrovic, S., Osborne, M., Lavrenko, V., 2010. Streaming

first story detection with application to twitter. In

Proceeding of HLT '10 Human Language

Technologies: The 2010 Annual Conference of the

North American Chapter of the Association for

Computational Linguistics.

Phuvipadawat, S., Murata, T., 2010. Breaking news

detection and tracking in twitter. In Web Intelligence

and Intelligent Agent Technology, IEEE/WIC/ACM

International Conference on, 3:120–123, 2010.

Qin, Y., Zhang, Y., Zhang, M., Zheng, D., 2013. Feature-

Rich Segment-Based News Event Detection on

Twitter. In International Joint Conference on Natural

Language Processing IJCNLP 2013.

Raad, EJ., Chbeir. R., 2014. Foto2Events: From Photos to

Event Discovery and Linking. In Online Social

Networks. SocialCom2014, Dec 2014, Sydney,

Australia.

Sakaki, T., Toriumi, F., and Matsuo, Y., 2011. Tweet

trend analysis in an emergency situation. In

Proceedings of the Special Workshop on Internet and

Disasters, SWID ’11, pages 3:1–3:8, New York, NY,

USA, 2011. ACM.

Srihari, R., Burhans, D., 1994. Visual Semantics:

Extracting Visual Information from Text

Accompanying Pictures. In Proceedings of AAAI-94.

Tov, EY., Diaz, F., 2011. Location and timeliness of

information sources during news events. In

Proceedings of SIGIR, pages 1105–1106, Beijing,

China.

Volkova, Q., Wilson, T., and Yarowsky, D., 2013.

Exploring Sentiment in Social Media: Bootstrapping

Subjectivity Clues from Multilingual Twitter Streams.

In Proc. ACL Vol2.

KDIR 2015 - 7th International Conference on Knowledge Discovery and Information Retrieval

300