=Paper=
{{Paper
|id=Vol-1329/papersmile_4
|storemode=property
|title=An Ontology-Based Approach to Social Media Mining for Crisis Management
|pdfUrl=https://ceur-ws.org/Vol-1329/papersmile_4.pdf
|volume=Vol-1329
|dblpUrl=https://dblp.org/rec/conf/esws/ZavarellaTSG14
}}
==An Ontology-Based Approach to Social Media Mining for Crisis Management==
https://ceur-ws.org/Vol-1329/papersmile_4.pdf
An Ontology-Based Approach to Social Media Mining
for Crisis Management
Vanni Zavarella, Hristo Tanev, Ralf Steinberger, and Erik Van der Goot
Joint Research Center - European Commission
Abstract. We describe an existing multilingual information extraction system
that automatically detects event information on disasters, conflicts and health
threats in near-real time from a continuous flow of on-line news articles. We illus-
trate a number of strategies for customizing the system to process social media
texts such as Twitter messages, which are currently seen as a crucial source of
information for crisis management applications. On one hand, we explore the
mapping of our domain model with standard ontologies in the field. On the other
hand, we show how the language resources of such a system can be built, start-
ing from an existing domain ontology and a text corpus, by deploying a semi-
supervised method for ontology lexicalization. As a result, event detection is
turned up into an ontology population process, where crowdsourced content is
automatically augmented with a shared structured representation, in accordance
with the Linked Open Data principles.
1 Introduction
Monitoring of open source data, such as news media, blogs or micro-blogging services,
is being considered worldwide by various security agencies and humanitarian organiza-
tions as an effective contribution to early detection and situation tracking of crisis and
mass emergencies. In particular, several techniques from text mining, machine learning
and computational linguistics are applied to user-generated (“crowdsourced”) content,
to help intelligence experts to manage the overflow of information transmitted through
the Internet, extract valuable, structured and actionable knowledge and increase their
situation awareness in disaster management[3].
A massive amount of messages on social media platforms such as Twitter, Face-
book, Ushahidi etc. are generated immediately after and during large disasters for ex-
change of real-time information about situation developments, by people on the affected
areas.The main added value of this content is that: a. it is nearly real-time, and typically
faster than mainstream news; b. it potentially contains fine-grained, factoid information
on the situation on the field, far more densely distributed geographically than official
channels from crisis management organizations. Nonetheless, there are several prob-
lems that hinder social media content to unfold its full potential for real-world applica-
tions.First, content is massive, containing a high rate of information duplication, partly
due to several people reporting about the same fact, partly due to platform-specific
content-linking practices, such as re-tweeting. Secondly, while social media messages
can be generated with a certain amount of platform-specific metadata (such as time,
56
geocoding and hashtags) a crucial part of their content is still encoded in natural lan-
guage text1 .
Both issues raise from the same general problem that crowdsourced content is un-
structured and lacks interpretation with respect to a shared conceptualization of the do-
main, so that it cannot be integrated with knowledge bases from humanitarian agency
information systems and thus be converted into actionable knowledge. This issue clearly
emerged in a survey with disaster management experts active on the field right after the
2010 Haiti earthquake, where Twitter users produced a significant amount of reports,
which were though underexploited because of the lack of semantic integration with ex-
isting information systems [3].One solution that has been explored consists in providing
the users with tools for structuring their content “on the fly”, for example by engaging
a layer of domain experts in encoding unstructured observations into RDF triples by
using the Ushahidi platform ([27]). We propose instead a general architecture for aug-
menting unstructured user contributions with structured data that are automatically ex-
tracted from the same user-generated content. The overall method is top-down.We then
apply a semi-supervised method for the lexicalization of the target ontology classes
and properties from text [1]. The method learns a mapping from classes of linguistic
constructions, such as noun and verb phrases, to semantic classes and event patterns,
respectively. Then, with a relatively limited human intervention, such constructions can
be linearly combined into finite-state grammars for detection of event reports. Finally,
we run the output grammar on crowdsourced content and populate the target ontology
with structured information from that content, by deploying an instance of the event
detection engine tuned to the social media streams.
As the ontology lexicalization method is language and domain independent, the
proposed architecture is highly portable across languages, including for instance the
il-formed ones used in social media platforms.
Next section is devoted to related work. Section 3 outlines the architecture of event
extraction from news text and describe its implicit domain model. Section 3 proceeds
by describing the ontology lexicalization algorithms. Section 5 shows how the result-
ing ontologies can be populated from user-generated content by information extraction
techniques. We conclude by exploring open issues and future developments of the pro-
posed architecture.
2 Related Work
Event detection and tracking in social media has gained increased attention in last years
and many approaches have been proposed, to overcome the information overload in
emergency management for large scale events, and to increase sensitivity of small case
incident monitoring ([20], [17],[29]). Different approaches vary with respect to the level
of analysis they perform on crowdsourced content and the amount of structure they
extract. This ranges from taking tweets as simple “sensor reading”, associated with a
1 The equally important issue of the confidence and trust of user-generated content, and the
specific techniques for content validation which can be deployed to cope with it, are out of the
scope of this paper.
57
time and location, of a target earthquake ([18]), to deriving RDF triples posts for content
merging ([19]).
Many studies have explored the use of ontologies for enhancing semantic interop-
erability in various domains of interest for the crisis management professionals (e.g.
spatial data and GIS, [21]). Nonetheless, a comprehensive standard ontology encom-
passing all those subject areas does not exist, while different sub-ontologies cover criti-
cal subject areas (such as Resource, Damage, Disaster) (see [6] for a survey). This poses
an issue of ontology mapping and integration for a crisis response information system
which would like to make use of Linked Open Data for its data sharing.
Methods for ontology learning and population have been explored in recent years
both within the Natural Language Processing community and in the Knowledge Repre-
sentation field (see [22] and [23] for an overview). Strongly related to our work is the
concept learning algorithm described in [24], which finds concepts as sets of semanti-
cally similar words by using distributional clustering.
Finally, there are many approaches for text mining from Twitter data ([25]). Rele-
vant to our approach are the methods for automatic event detection from Twitter like the
one described in [26]. However, in contrast to the already existing approaches, we per-
form structured event extraction from the tweets rather than only event detection-based
tweet classification.
3 Event Extraction Architecture
We have created a multilingual event extraction engine NEXUS for global news mon-
itoring of security threats, mass emergencies and disease outbreaks [9]. The system is
part of the Europe Media Monitor family of applications (EMM) [8], a multilingual
news gathering and analysis system which gathers an average of 175,000 online news
articles per day in up to 75 languages, taken by a manually selected set of mainstream
newspapers and news portals. NEXUS builds upon the output of EMM modules and
identifies violent events, man-made and natural disasters and humanitarian crises from
news reports. It then fills an event type-specific template like the one depicted in Figure
1, including fields for number and description of dead, injured, kidnapped and displaced
people, descriptions of event perpetrators, event time and location, used weapons, etc.
Nexus can work in two alternative modes, cluster-based and on single articles. In
the latter the full article text is analysed. In the former, we process only the title and the
first three sentences, where the main facts are typically summarized in simple syntax
according to the so-called “inverted pyramid” style of news reports. 2 .
Figure 2 sketches the entire event extraction processing chain, in cluster-based
mode.
First, news article are scanned by EMM modules in order to identify and enrich
the text representation with meta-data such as entities and locations, that are typically
2 This is feasible without significant loss in coverage as the news clusters contain reports from
different news sources about the same fact, so that this redundancy mitigates the impact on sys-
tem performance of linguistic phenomena which are hard to tackle, such as anaphora, ellipsis
and long distance dependencies.
News
News
News
News
News
Meta-Data Creation
Real Time Clusterer
Geo-Locating
Core Linguistic Analysis
Cascaded Event
Extraction Grammar
Application
Entity Role Victim Event Type
Disambiguator Arithmetic Classifier
Event Description
Assembly
Fig. 1. The output structure of the Fig. 2. Event extraction processing chain
event extraction system
separate from the ones deployed in the event extraction process proper. Next the arti-
cles are clustered and then geo-located according to extracted meta-data. Each article in
the cluster is then linguistically preprocessed by performing fine-grained tokenization,
sentence splitting, domain-specific dictionary look-up (i.e. matching of key terms indi-
cating numbers, quantifiers, person titles, unnamed person groups like “civilians”, “po-
licemen” and “Shiite”), and finally morphological analysis, simply consisting of lexicon
look-up on large domain-independent morphological dictionaries. The aforementioned
tasks are accomplished by CORLEONE (Core Linguistic Entity Online Extraction), our
in-house core linguistic engine.
Subsequently, a multi-layer cascade of finite-state extraction grammars in the Ex-
PRESS formalism [12] is applied on such more abstract representation of the article
text, in order to: a) identify entity referring phrases, such as persons, person groups, or-
ganizations, weapons etc. b) assign them to event specific roles by linear combination
with event triggering surface patterns. For example, in the text “Iraqi policemen shot
dead an alleged suicide bomber” the grammar should extract the phrase “Iraqi police-
men” and assign to it the semantic role Perpetrator, while the phrase “alleged suicide
bomber” should be extracted as Dead.
We use a lexicon of 1/2-slot surface patterns of the form:
was shot by
has been taken hostage
where each slot position is assigned an event-specific semantic role and includes a type
restriction (e.g. Person) on the entity which may fill the slot. ExPRESS grammar rules
are pattern-action rules where the left-hand side (LHS) is a regular expression over Flat
59
Feature Structure (FFS) and the right-hand side (RHS) consists of a list of FFS, which
is returned in case the LHS pattern is matched. See Section 5 for a sample rule.
The systems includes an event type classification module consisting of a blend of
keyword matching, event role detection and a set of rules controlling their interaction.
First, for each event type, we deploy: a) a list of weighted regular expression keyword
patterns, allowing multiple tokens and wild cards b) a set of boolean pattern combina-
tions: OR pattern lists are combined by the AND operator, each pattern is a restricted
regular expression and conjunctions are restricted by proximity constraints.
As contradictory information on the same event may occur at the cluster level, a last
processing step in the cluster-based mode consists of cross-article cluster-level infor-
mation fusion: that is, the system aggregates and validate information extracted locally
from each single article in the same cluster, such as entity role assignment, victim counts
and event type.
The system architecture is highly customizable across languages and domains, by
making use of simple local parsing grammars for semantic entities, backed by a num-
ber of lexical resources learned by semi-supervised machine learning methods [11].
Currently instances are in place for English, French, Italian, Spanish, Portuguese, Ro-
manian, Bulgarian, Czeck, Turkish, Russian and Arabic language. The live event extrac-
tion results are freely accessible online for anybody to use, both in text form (http://
emm.newsbrief.eu/NewsBrief/eventedition/all/latest.html) and
displayed on a map (http://emm.newsbrief.eu/geo?format=html&type=
event&language=all).
While extracted data are not delivered by the system in any Linked Data standard, a
fine-grained categorization of domain entities and relations is implicitly used through-
out the extraction process. An ontology representation of the domain model shared by
the different instances of Nexus engine is illustrated in Figure 3.
As it can be seen, currently the system is profiled to model crisis event occurrence,
rather than emergency tracking and relief operations. However, we experimented with
a statistical method for building semi-automatically the core resources of an event ex-
traction engine for a given language, starting from a pre-existing ontology and a text
corpus.
4 Ontology Lexicalization Method
The general schema of our method is outlined here (see [1] for more details). Given a
set of classes and properties from an event ontology:
1. we learn terms which refer to ontology classes by using a multilingual semantic
class learning algorithm. These classes typically represent event-related entities,
such as Building and EmergencyCrew.
2. we learn pre-and post-modifiers for the event-related entities in order to recognize
entire phrases describing these entities. We do not give details on this part here (see
[1]).
3. we learn surface event-triggering patterns for the properties relating ontology classes.
As an example, the pattern [BUILDING] was destroyed instantiate the property in-
volvesBuilding relating the event class BuildingDamage to a Building entity.
Fig. 3. Ontology representation of the Event Extraction system domain model.
4. Using the lexical classes and patterns learned in steps 1, 2 and 3, we create a finite-
state grammar to detect event reports and extract the entities participating in the
reported events, using the text processing modules and grammar engine described
in Section 3.
4.1 Semantic Class Learning Algorithm
The algorithm we describe here accepts as input a list of small seed sets of terms, one for
each semantic class under consideration in addition to an unannotated text corpus. Then,
it learns additional terms that are likely to belong to each of the input semantic classes.
As an example for English language, starting with two semantic classes Building and
Vehicle, and providing the seed term set home, house, houses and shop, and the seed
term set bus, train and truck, respectively, the algorithm will return extended classes
which contain additional terms like cottage, mosque, property, for Building and taxi,
61
lorry, minibus, boat for Vehicle. The semantic class learning algorithm looks like the
following:
input : OriginalSeedSets- a set of seed sets of words for each considered
class; Corpus- non-annotated text corpus; NumberIterations-
number of the bootstrapping iterations
output : Expanded semantic classes
CurrentSets OriginalSeedSets;
for i 1 to NumberIterations do
CurrentSets SeedSetExpansion(CurrentSets, Corpus) ;
CurrentSets ClusterBasedTermSelection(CurrentSets,
OriginalSeedSets, Corpus) ;
end
return CurrentSets
It makes use of two sub-algorithms for: (a) Seed set expansion; (b) Cluster-based
term selection.
The seed set expansion learns new terms which have similar distributional features
to the words in the seed set. It consists of two steps: (a)Finding contextual features
where, for each semantic class ci , we consider as a contextual feature each uni-gram
or bi-gram n that co-occurs at least 3 times in the corpus with any of its seed terms
seed(ci ) and that is not composed only of stop words (we have co-occurrence only
when n is adjacent to a seed term on the left or on the right); (b) Extracting new terms
which co-occur with the extracted contextual features. In other terms, we take for each
category the top scored features and merge them into a contextual-feature pool, which
constitutes a semantic space where the categories are represented, and represent each
term t as a vector vcontext (t) in the space of contextual features. Then we score the
relevance of a candidate term t for a category c by using the projection of the term
vector on the category vector.
The presented algorithm is semi-supervised ([16]) and in order to improve its pre-
cision we introduce a second term selection procedure: the learned terms and the seed
terms are clustered, based on their distributional similarity; then, we consider only the
terms which appear in a cluster, where at least one seed term is present. We call these
clusters good clusters. The increased precision introduced by the cluster-based term
selection allows for introducing bootstrapping in our process, which otherwise is typi-
cally affected by the problem of error propagation across iterations (so called “semantic
drifting”).
4.2 Learning of Patterns
The pattern learning algorithm acquires in a weakly supervised manner a list of pat-
terns which describe certain actions or situations. We use these patterns to detect event
reports.
Each pattern looks like the ones shown in Section 3. For example, the pattern dam-
aged a [BUILDING] will match phrases like damaged a house and damaged a primary
school.
62
The algorithm accepts as its input: (a) A list of action words, e.g. damaged, dam-
aging, etc. (b) a representation of the semantic category for the slot as a term list, e.g.
house, town hall, etc. (c) an unannotated text corpus. It then returns a list of additional
patterns like [BUILDING] was destroyed.
The main idea of the algorithm is to find patterns which are semantically related
to the action, specified through the input set of action words and at the same time will
co-occur with words which belong to the semantic class of the slot. It consists of three
steps (see [1]):
1. find terms similar to the list of action words, e.g. destroyed, inflicted damage, using
the semantic class expansion algorithm above;
2. learn pattern candidates which co-occur with the slot semantic category (e.g. Build-
ing), using the contextual feature extraction sub-algorithm. Each contextual feature
of the slot class is considered a candidate pattern.
3. select only candidate patterns which contain terms similar to the action words (dis-
covered in the first step). In this way, only contextual patterns like inflicted damage
on a [BUILDING] will be left.
The algorithms presented here make use of no language analysis or domain knowl-
edge, consequently they are applicable across languages and domains. In particular,
this makes our method potentially applicable to the ill-formed language used in Twitter
and other social media, provided that a relatively large text corpus is available.
5 Populating Disaster Management Ontologies
Once the lexicalization of an ontology has been carried out in a target language, we can
apply part of the event extraction infrastructure described in Section 3 to populate that
ontology from text streams. In [1] we report about an evaluation on populating a micro-
ontology for disaster management with instances of events extracted from a stream of
tweets published during several big tropical storms, in English and Spanish language.
The ontology structure is shown in Figure 4
Fig. 4. The micro ontology used for the experiment
Each object from the sub-types of Event, namely BuildingDamage, InterruptionOfC-
ityService, RestaurationOfCityService and DeploymentOfEmergencyCrew, is related to
63
exactly one object from a sub-class of ParticipatingEntity via has-a-participant rela-
tion. Consequently, the detection of an event report can be done by detecting an action
or situation with one participant. Event detection was performed by devising a sim-
ple two-level grammar cascade, whose rules combine semantic classes, patterns and
modifiers learned with the previously described algorithms. Namely, it detects at the
first level event participant entities: then, it combines them with event patterns to detect
event-specific actions or situation, by rules like the one in Figure 5: where the boolean
event_rule :> ( leftPattern & [CLASS:"A"]
participatingEntity & [CLASS:"C", SURFACE:#surf] )
-> event & [CLASS:"A", PARTICIPANT:#surf ]
& PossibleSlotFor(A,C).
Fig. 5. Rule schema for single participant event detection.
operator PossibleSlotFor(A) returns a list of all possible sub-classes of Participatin-
gEntity which may fill the slots of a pattern of class A. As an example, when matching
an BuildingDamage pattern followed by a Building entity expression on a tweet like
@beerman1991 yeah...i mean, there was thatb neighborhood
in queens where like, 70 houses burned down during #sandy
this rule will populate the micro-ontology with an instance of the BuildingDamage
event class and an instance of Building class, and relate them through an instance of
has-a-participant property.
On an evaluation performed on a mixed-language English and Spanish corpus of
270,000 tweets, this method proved accurate in extracting events (87% and 95% for
English and Spanish, respectively), while recall on a small test set turned up to be very
low (23% and 15%). This seems to be due to missing action/situation patterns and to
the finite-state grammars not being able to parse pattern-entity combination.
We plan to tackle the current limitation by deploying the ontology-derived grammar
within the full-fledged event extraction infrastructure outlined in Section 3, where bag-
of-words methods would be put into place, parallel to strict grammar matching, for the
instantiation of event type entities and the extraction of participant entities.
To this end, we sketch a general strategy to map the current event extraction en-
gine to the information structure of existing ontologies for Crisis Management. First,
we manually map the implicit data model outlined in Section 3 with domain ontologies
such as MOAC (http://observedchange.com/moac/ns#) or HXL (http:
//hxl.humanitarianresponse.info/ns/) and IDEA (Integrated Data for
Events Analysis, see http://vranet.com/IDEA.aspx). We believe that, by do-
ing this, we would increase the usefulness of our system output, both for human users
working in the field of crisis management or conflict resolution and for services making
use of Linked Open Data resources. In some cases, a one-to-one mapping already ex-
ists, as for instance between the hxl:PopulationGroup and nexus:PersonGroup classes,
64
or between moac:Floods and nexus:Floods. In other cases, our ontology event type en-
tities constitute an application of the more generic hxl:Incident entity.
Then, we add the missing, emergency-related sub-ontologies, encompassing classes
like moac:InfrastructureDamage, moac:ContaminatedWater and moac:PowerOutag and
we customize the system to detect events from them, via the method described in the
previous section. Figure 6 illustrates the whole process:
Fig. 6. A general schema for customizing the ontology-based event extraction engine.
6 Summary and Future Work
We outlined a procedure for mapping the domain model of an existing event extrac-
tion engine to larger scope ontologies for Crisis Management, in order to perform
ontology-based event detection on social media streams. While the core of the proce-
dure is a language-independent ontology lexicalization method that proved promising
in supporting text mining from social media, processing of such sources remains chal-
lenging because: (a) potentially relevant messages have to be filtered from the large
amount of user messages and (b) the usage of ill-formed text in social media messages
(lower-casing names, omitting diacritics, doubling letters for stress etc.) reduces the
information extraction recall and accuracy of automatic systems. We partially address
(a) by generating a keyword-based query to retrieve event-related tweets starting from
the text of a news cluster about that event ([28]). This will only be applicable though
to large events which make to news clusters, while for the small scale incidents typi-
cally targeted for emergency tracking some methods for message duplication detection,
clustering and merging appear necessary. (b) is currently addressed by deploying a pre-
65
processing layer for tweet language normalisation ([13]) but work is still to be done in
this direction.
Finally, as a prospective evaluation exercise, we plan to use a test corpus of around
18000 SMS messages (manually translated into English) collected by the Ushahidi plat-
form, in the context of the Mission 4636 relief project, soon after the Haiti 2010 earth-
quake. We are currently searching for some validation event data from that context for
measuring the extractive performance of our system.
References
1. Tanev, H. and Zavarella, V. (under publication) ‘Multilingual Lexicalisation and Population
of Event Ontologies. A Case Study for Social Media’.
2. Ashish, N., Appelt, D., Freitag, D. and Zelenko, D. (2006) ‘Proceedings of the workshop on
Event Extraction and Synthesis’, held in conjunction with the AAAI 2006 conference, Menlo
Park, California, USA.
3. Ortmann, J., Limbu, M., Wang, D., Kauppinen, T. (2011) ‘Crowdsourcing Linked Open Data
for Disaster Management’, Proceedings of the 10th International Semantic Web Conference,
Germany, 11-22.
4. Fan, Z., Zlatanova, S. (2011) ‘Exploring Ontologies for Semantic Interoperability of Data in
Emergency Response’, Appl Geomat, 3, 2, 109-122.
5. Gruber, T. (2008) ‘Collective knowledge systems: Where the Social Web meets the Semantic
Web’, Web Semantics: Science, Services and Agents on the World Wide Web, Volume 6
Number 1, pp. 4-13.
6. Shuangyan Liu and Duncan Shaw and Christopher Brewster (2013) ‘Ontologies for Crisis
Management: A Review of State of the Art in Ontology Design and Usability’, Proceedings
of the Information Systems for Crisis Response and Management conference (ISCRAM 2013
12-15 May, 2013)
7. Han, Bo, Cook, P. and Baldwin, T. (2012)‘Automatically Constructing a Normalisation Dic-
tionary for Microblogs’, Proceedings of the 2012 Joint Conference on Empirical Methods
in Natural Language Processing and Computational Natural Language Learning, EMNLP-
CoNLL ’12, Jeju Island, Korea, pp. 421–432.
8. Steinberger, R., B. Pouliquen, and E. van der Goot. (2009). ‘An introduction to the europe
media monitor family of applications’. In Information Access in a Multilingual World - Pro-
ceedings of the SIGIR 2009 Workshop, Boston, USA.
9. Piskorski, J. and Tanev, H. and Atkinson, M., Van der Goot, E. and Zavarella, V. (2011) ‘On-
line News Event Extraction for Global Crisis Surveillance’, Lecture Notes in Computer Sci-
ence, volume 6910, pp. 182-212.
10. Tanev, H., Piskorski, J., Atkinson, M. (2008) ‘Real-Time News Event Extraction for Global
Crisis Monitoring’, NLDB ’08 Proceedings of the 13th international conference on Natural
Language and Information Systems: Applications of Natural Language to Information Sys-
tems, pp. 207 - 218.
11. Tanev, H., Zavarella, V. Linge, J. Kabadjov, M., Piskorski, J. Atkinson, M. and Steinberger,
R.. (2009). ‘Exploiting Machine Learning Techniques to Build an Event Extraction System
for Portuguese and Spanish’, Linguamatica, 1(2):5566.
12. Piskorski, J. (2007). ‘ExPRESSextraction pattern recognition engine and specification suite’.
In Proceedings of the International Workshop Finite-State Methods and Natural language Pro-
cessing.
13. Küçük D., Guillaume J. and Steinberger, R. (2014) ‘Named Entity Recognition on Turkish
Tweets’, Proceedings of the Conference Language Resources and Evaluation’ (LREC2014).
66
14. Xiaohua Liu, Ming Zhou, Furu Wei, Zhongyang Fu, and Xiangyang Zhou. (2012). ‘Joint
Inference of Named Entity Recognition and Normalization for Tweets’. In Proceedings of the
50th Annual Meeting of the Association for Computational Linguistics, pp. 526535.
15. OBrien Sean P. (2002). ‘Anticipating the Good, the Bad, and the Ugly: An Early Warning
Approach to Conflict and Instability’. Journal of Conflict Resolution 46:791.
16. Tanev,H. and Magnini,B. (2008). ‘Weakly Supervised Approaches for Ontology Population’.
Ontology learning and population. Bridging the gap between text and knowledge, Berlin:
Springer Verlag.
17. Schulz, A. Ristoski, P., Paulheim, H. (2013) ‘I See a Car Crash: Real-Time Detection of
Small Scale Incidents in Microblogs’ Lecture Notes in Computer Science Volume 7955, 2013,
pp 22-33.
18. Sakaki, Takeshi and Okazaki, Makoto and Matsuo, Yutaka (2010) ‘Earthquake Shakes Twit-
ter Users: Real-time Event Detection by Social Sensors’, Proceedings of the 19th International
Conference on World Wide Web, ACM, pp 851–860.
19. Schulz, A. and Ortmann, J. and Probst, F. (2012) ‘Getting User-Generated Content Struc-
tured: Overcoming Information Overload in Emergency Management’, Global Humanitarian
Technology Conference (GHTC), 2012 IEEE, pp 143-148.
20. Vieweg, S. and Hughes, A.L. and Starbird, K. and Palen, L., (2010) ‘Microblogging Dur-
ing Two Natural Hazards Events: What Twitter May Contribute to Situational Awareness’
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, ACM, pp
1079-1088.
21. Guarino, N. (1998) ‘Formal Ontology and Information Systems’ in FOIS 2004: Proceedings
of the Third International Conference. IOS Press pp 315.
22. Drumond,L. and Girardi,G. (2008) ‘A survey of ontology learning procedures’. In The 3rd
Workshop on Ontologies and their Applications, pp.1325, Salvador, Brazil.
23. Buitelaar, P., and Cimiano, P. (eds.) (2008) ‘Ontology learning and population. Bridging the
gap between text and knowledge’, Berlin: Springer Verlag.
24. Pantel,P. and Lin, D. (2002) ‘Discovering Word Senses from Text’. In Proceedings of ACM
SIGKDD Conference on Knowledge Discovery and Data Mining. pp. 613619. Edmonton.
25. Breslin, J., Ellison, N., Shanahan, J., and Tufekci, Z. (Eds.) (2012) ‘Proceedings of the 6th
International Conference on Weblogs and Social Media (ICWSM - 12)’, Dublin: AAAI Press.
26. Reuter,T. and Cimiano, P. (2002) ‘Event-based classification of social media streams’. In
Proceedings of the 2nd ACM International Conference on Multimedia Retrieval, Hong Kong.
27. Okolloh, O. (2009) ‘Ushahidi, or testimony: Web 2.0 tools for crowdsourcing crisis informa-
tion’, Participatory Learning and Action, vol. 59, no. 1, pp. 6570.
28. Tanev H., Ehrmann M. Piskorski, J. and Zavarella, V. (2012) ‘Enhancing Event Descriptions
through Twitter Mining’. In: AAAI Publications, Sixth International AAAI Conference on
Weblogs and Social Media, pp 587-590.
29. Ritter, Alan and Mausam and Etzioni, Oren and Clark, Sam (2012) ‘Open Domain Event
Extraction from Twitter’. In Proceedings of the 18th ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining,KDD ’12, Beijing, China, pp. 1104–1112.