Exploring the Role of Named Entities for Uncertainty Recognition in
Event Detection
Masnizah Mohd and Kiyoaki Shirai
Japan Advanced Institute of Science and Technology, 1-1 Asahidai, Nomi, Ishikawa, 923-1292, Japan
Keywords: Uncertainty, Named Entities, User, Event Detection.
Abstract: Ambiguous information contributes to the uncertainty issue. Type of information such as using named entities
has been proved to provide significant information to the user compared to the ‘bag-of-words’ in identifying
an event. So what else could contribute to the uncertainty in an event detection? We propose to answer this
question by analysing the distribution of named entities across topics, and explore the potential of named
entities in a user experiment. We construct an event detection task with 20 users and use news dataset from
Topic Detection and Tracking (TDT) corpus, under the Sports and Politics categories. We analyse the results
from five uncertainty dimensions: too little information, too much information, complex information,
ambiguous information and conflicting information. These dimensions are categorise as two factors; amount
and type of information. There was no statistical significance difference in the amount of information given
with the number of successful event detected. However, with little information and high named entities has
contributes in reducing uncertainty. In addition, the amount of information and information quality are
mutually independent. Our results suggest that uncertainty vary substantially between the amount of
information and type of information in event detection.
1 INTRODUCTION
“Uncertainty” in event detection and tracking task can
be interpreted as lack of information and inability to
interpret or determine an event due to the little or too
much information. This task rely on named entities as
one of the important features used to detect an event
which occurs in a topic.
The objective of this experiment is to explore the
potential of named entities (NEs) for uncertainty in
event detection. Therefore we take two approaches.
The former, proposed by Mohd and Mabrook (2014)
in the context of Topic Detection and Tracking (TDT)
systems, proved that named entities was useful in
improving Tracking task (including sub Profiling
activity) performance meanwhile bag of words
(BOW) improved user performance in Detection task.
Therefore type of information either NEs or BOW
were considered. However the potential role of NEs
in reducing or increasing uncertainty in TDT has
never been explored.
In the second approach, we go beyond ‘type of
information’ by also considering the ‘amount of
information or stories’ provided. Previously, Hurley
et al., (2011) evaluated uncertainty in online news
focusing on cancer topic from 5 dimensions; too little
information, too much information, ambiguous
information, complex information and conflicting
information. Therefore in this experiment we
considered the 5 dimensions of uncertainty
introduced by Hurley et al., (2011) with the type of
information. We want to investigate the potential role
of NEs or BOW for uncertainty recognition in event
detection.
These two approaches are important to validate
research questions:
Is there any relationship between the ‘type of
information’ and ‘amount of information’ for
uncertainty recognition in event detection?
What is the ‘type of information’ that are
considered as complex, ambiguous and
conflicting to the user?
We conducted a set of user experiments that concern
various kinds of entities (e.g. Person, Location,
Organization, Date, Time, Money, and Percent)
across topics (Politic, Sports). Our experiment
include 1000 evaluations of news stories.
This paper is organized as follows. Section 2
describes related work and positions our approach.
Mohd, M. and Shirai, K..
Exploring the Role of Named Entities for Uncertainty Recognition in Event Detection.
In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 3: KMIS, pages 335-341
ISBN: 978-989-758-158-8
Copyright
c
2015 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
335
Section 3 discuss the methodology applied to
construct the user experiment. Finally in Section 4,
we report the findings. We end the paper with
conclusions and thoughts for future work.
2 RELATED WORK
Uncertainty is one of the challenges in information
seeking and retrieval (Chowdhury et al., 2011). Many
attempts have been done in developing uncertainty
model by investigating human information behaviour
in information seeking and retrieval process
(Ingwersen, 1992). There are few work that proposed
natural language processing technique such as from
syntactical and semantic approach to reduce
uncertainty (Goodman, 2008; Topka, 2013).
Several linguistic research aim at modelling the
use of modality, but very few concentrate on
uncertainty, for instance the Certainty Categorization
Model proposed by Rubin (2006). This model was
based on four dimension; Level, Perspective, Focus
and Time to characterize uncertainty. For level
dimension, they considered the words such as ‘might
buy’ and ‘will come’ to be classified as Absolute level
or Low level. Meanwhile in Perspective level, they
analysed on how the sentences are reported from
writer’s point of view. Focus dimension differentiated
between Abstract and Factual information. Finally for
Time dimension, they analysed the sentences based
on past, present and future time. Then Goujon (2009)
enhanced the Certainty Categorization Model
proposed. The enhanced model includes the
identification of the local source, which was
important to the end user in validating the reliability
of the reported discourse. It also takes into account
the reality and unreality of an information which was
specified in the source text, rather than the Focus
dimension. Thus the enhanced dimensions consist of
five; Level, Perspective, Time, Reality and Source
Name to characterize uncertainty.
There are also few work in measuring uncertainty
in message. Mishel (1988) has introduced forms of
uncertainty (ambiguity, complexity, volume of
information and unpredictability) and Babrow (1998)
dimensions of uncertainty were combined to form
five forms of dimensions of uncertainty in messages.
Instances of uncertainty related content within a
message are such as message characteristic (specific
words, phrases or sentences). Then Hurley (2011)
enhanced the dimension of uncertainty into five
dimensions: too little information (volume), too much
information (volume), complex information,
ambiguous information and conflicting information.
These five forms of uncertainty in messages was
easily identified in news article and been
implemented in cancer news article.
In the context of TDT research, researchers have
attempted to build better document models,
developing similarity metrics or better document
representations (Chen and Ku, 2002). This has led to
a series of research efforts that concentrate on
improving document representation by applying
Named Entity Recognition (Chen and Ku, 2002).
Mohd and Mabrook (2014) investigated the potential
of named entities in TDT tasks and they discovered
that NEs has improved both tasks. However there is
no work has evaluate the role of NEs for uncertainty
recognition in event detection task. This is the first
work that explored the five dimensions of uncertainty
in TDT.
3 METHOD
There are two approaches in this work. First we
analysed the distribution of named entities (NEs)
across topics (Section 3.2) and secondly we
conducted a user experiment (Section 3.3 - 3.4) to
explore the potential of named entities for uncertainty
recognition in event detection task.
3.1 Dataset
We used 300 news documents from Topic Detection
and Tracking (TDT) corpus. There are 2 categories
(Politics and Sports) with 10 topics and 50 events
occurred as shown in Table 1. On average, there are
5 events and 30 documents/story per topic. In TDT, a
topic consist of several events and an event consist of
several stories or documents.
Table 1: Topics and events for Politics and Sports
categories.
Topic: [P1] Current Conflict with Iraq (20015)
Event
Current Conflict with Iraq
Iraq announces it will block inspections
Iraq prevents inspection team from entering
Reaction to blocked inspection team
Inspection team withdrawn
Hussein may stop cooperating with inspections
Topic: [P2] Clinton-Jiang Debate (20096)
Event
Plans, preparations for Clinton's trip to China
Clinton leaves for China
Clinton's activities in China
Freedom of worship for Chinese citizens
Reaction to Clinton's trip
RDBPM 2015 - Special Session on Research and Development on Business Process Management
336
Table 1: Topics and events for Politics and Sports
categories. (cont.)
Topic: [P3] Gingrich Resigns (30024)
Event
Reaction to elections, Gingrich faces challenge to
speakership
Largent, Livingston to challenge GOP leaders
Gingrich announces he will resign
Reaction to, reflection on Gingrich resignation
Candidates emerge for speakership
Topic: [P4] US Mid-term Elections (30050)
Event
Clinton campaigns for Democrats
Impeachment hearings begin
Effect of impeachment hearings on campaigns
Budget negotiations
Effect of budget on campaigns
Impact of other issues on campaigns
Topic: [P5] Clinton's Gaza Trip (30053)
Event
Clinton visits Middle East
Police fire on West Bank demonstrators
White House praises PLO revocation
Clinton comments on impeachment hearings
Clinton meets with Middle East leaders
Netanyahu will not hand over more land
Topic: [S1] 1998 Winter Olympics (20013)
Event
Preparation for Olympics
Olympic games open
Olympic contests, results
Topic: [S2] Super Bowl '98 (20033)
Event
Preparations, predictions for Super Bowl
Broncos win Super Bowl
Post-game celebrations, riots
Topic: [S3] NBA finals (20087)
Event
Basketball regional finals
Finals
Bulls win championship, Chicago celebrates
Topic: [S4] Yankees vs. Padres in World Series
(31026)
Event
Padres win NLCS
Yankees win ALCS
Game 1 of World Series
Joe DiMaggio, Darryl Strawberry illnesses
Game 2 of World Series
Game 3 of World Series
Topic: [S5] Joe DiMaggio Illness (31036)
Event
DiMaggio in hospital for pneumonia
Debate, discussion over heart attack and lung cancer
Doctors confirm DiMaggio had lung cancer
Reflection on DiMaggio
DiMaggio develops infection, improves, then coma
DiMaggio improves
DiMaggio tells doctors to stop updating press
3.2 Named Entity Recognition
We used ANNIE (A Nearly-New Information
Extraction System) that has been developed using
GATE (Cunningham, 2002). It is an example of a
lexical resource and rule-based approach to IE. It was
used to identify regions of text corresponding to the
seven MUC-7 named entity types (Person, Location,
Organization, Date, Time, Money, and Percent).
The ANNIE system consists of seven processing
resources organized into an application pipeline.
These include a tokenizer, a gazetteer, a sentence
splitter, a POS (parts of speech) tagger, and a named
entity transducer. Each of them is associated with a
language resource containing data or rules, i.e.
tokenizer rules, gazetteer lookup lists, sentence
segmentation rules, a POS lexicon, and NE
transduction rules. The resources that are rule-based
use JAPE (Java Annotations Pattern Engine)
grammar rules to match patterns using regular
expressions over annotations in order to create new
annotations. JAPE rules can match against
annotations, annotation features, token attributes,
lookup types, and/or parts of speech and can take any
java-based action in response to a matched pattern.
Based on ANNIE’s capability, therefore we were not
building a NER system and instead using the existing
system to recognise named entities in a document. We
used it for its accurate entity, pronoun and nominal
co-references extraction.
3.3 Procedure
We conducted a user experiment with 20 users from
October, 2014 to January, 2015 at the School of
Information Science, Japan Advanced Institute of
Science and Technology (JAIST). The users were
postgraduate students and the average age of the users
was 20–30 years. Users were asked about their topic
familiarity and topic interest before they started with
the task. 1000 tasks were performed (20 users, 10
topics, 5 events) in this experiment. Five topics on
Politic (P1-P5) and five topics on Sports (S1-S5) were
conducted in two sessions in the Event Detection task
(Section 3.4). After completion of the tasks, users
were interviewed about their opinion or experience in
performing the task. They were given 15 minutes to
identify an event for each topic. The time assigned to
each task was sufficient based on the feedback
received from the pilot test conducted. The users were
offered a short break (5–15 minutes) after the first
session. A Latin square (Sparck-Jones, 1997) was
used to construct the experimental design (refer Table
2). This allowed us to evaluate the same topic using
Exploring the Role of Named Entities for Uncertainty Recognition in Event Detection
337
different amount of information. This was important
to justify whether the type of information helped the
users to detect an event even though they were given
a low amount of information.
Table 2: Experimental design.
Users
Session 1 Session 2
Topic (Politic) Topic (Sports)
P1 P2 P3 P4 P5 S1 S2 S3 S4 S5
1-10 L L L L L H H H H H
11-20 H H H H H L L L L L
L=Too little information/stories (Low)
H=Too much information/stories (High)
Low information means users received 40% or
less stories while High information means they
received 70% or more stories. 12 stories will be
selected from the 30 stories using random sampling
method to unsure equal probability of selection for
each article in an event under the Low information
category.
We also gave attention to the information quality.
The ability to detect an event might be associated with
the amount and quality of information given. User
might claimed that they received poor information,
hence we would like to avoid any issue due to the
quality of information provided during the
experiment. Thus these 2 aspects; amount and
information quality are mutually independent. The
stories in TDT corpus consist of quality information
as it reflect the way it was annotated (TDT
Annotation Manual). An integral and key part of the
corpus is the annotation of the corpus in terms of the
events discussed in the stories as shown in Table 3.
Table 3: Example of story for topic ‘Current Conflict with
Iraq’.
Topic Event Story
Current
Conflict
with Iraq
Iraq
announces it
will block
inspections
Iraq says it will block one of the
U.N. inspection teams being led
by an American until the team is
recomposed with fewer
Americans. The team carried out
its duties today without
interference.
Iraq
prevents
inspection
team from
entering
A new crisis may be developing
in Baghdad. The Iraqi
government blocked an
American-led team of U.N.
weapons inspectors from doing its
work today. The White House
says it's coordinating a response at
the U.N. Security Council.
We defined successful event based on the
keywords or the important terms in the event detected
given by user. We classified event into three
categories:
a. None: where users did not provide any event or
they did not complete the task;
b. Successful: where users provide the right
keyword or the important terms in the event
detected
c. Unsuccessful: where users failed to provide the
right keyword or the important terms in the event
detected
We also conflated different keywords referring to the
same context and meaning as shown in Table 4.
Table 4: Example of event detected by users and their
category.
Predefined
event
Event detected by users Category
Gingrich
announces he
will resign
Gingrich resignation Successful
Newt Gingrich leave job Successful
Newt Gingrich
dissatisfaction
Unsuccessful
Gerald Ford resign Unsuccessful
3.4 Event Detection Task
The event detection task was designed based on five
uncertainty dimension. A user’s session consisted of
the following stages, carried out in a single block of
time. In this task, the users had to detect an event for
a given topic. The procedure for performing the task
was as follows.
a. Users were welcomed and asked to read the
introduction to the experiment provided on an
information sheet. This set of instructions was
developed to ensure that each user received
precisely the same information. Users could retain
the information sheet after the user experiment.
b. The users were given a short overview of what the
experiment would entail. We also explained our
role in this experiment – i.e. to provide users with
support and remind users of the time taken in
performing the task.
c. Users were asked to complete an entry
questionnaire to provide us with background
information.
d. Event Detection task
Users were asked to perform the task by
identifying what was the event by following
the experimental design (as shown in Table 2).
Users were given 15 minutes to identify an
event, and could stop early if they were unable
to find any more relevant information.
Then there was one sub-activity in this task:
profiling. Profiling required the user to
provide the important keyword that was
RDBPM 2015 - Special Session on Research and Development on Business Process Management
338
considered as; ambiguous, complex and
conflicting information (refer Table 5).
Table 5: Definition of ambiguous, complex and conflicting
information.
Category Definition
Ambiguous
Keyword which is not clear and have
several possible meanings or
interpretations in detecting an event.
Complex
Keyword which is difficult to understand
in detecting an event.
Conflicting
Keyword which is contradict or different
in detecting an event.
e.
Users performed the post-evaluation interview. In
this session, we asked user about their experiences
in performing the task.
4 RESULTS
Findings revealed that there was no statistical
significance difference between topics and topic
familiarity (Mann-Whitney Test, p=0.496). The users
were not familiar with the topics given in the Event
Detection task (mean=2.04 sd=1.05). There were also
no statistically significant difference between the
users and their topic interest (Mann-Whitney Test,
p=0.844). Their topic interest was average
(mean=3.29 sd=1.11). This is a good indication of the
experiment since the users are not affected by
external factors such as their topic familiarity and
topic interest.
4.1 Named Entity Distribution Across
Topics
Figure 1: Named entities across topics.
Figure 1 summarizes the distribution of NE across
topics. Topic P5 (Clinton's Gaza Trip) has the highest
percentage of Person (19.4%), Money (14.3%) and
Percent (13.0%) NEs. Meanwhile topic S4 (Yankees
vs. Padres in World Series) has the highest
Organisation (18.1%) and Date (17%) NEs. Topic P3
(Gingrich Resigns) has the highest Location NEs
(20.2%). The distribution of NEs are affected by the
topics and events occurred. One possibility is the
nature of the topic that has caused certain NEs to
appear frequently.
4.2 User Evaluation
In this section we discussed the evaluation on the
amount and type of information by analysing the rate
of successful event detected (discussed in Section
3.3) in conjunction with the amount of
information/stories and named entities distribution
across topics.
4.2.1 Amount of Information
The entire event detection task was successful, with
93.9% of the task being successful and 6.1% being
unsuccessful.
Figure 2: Successful event detection rate for high and low
information/stories across topics.
We associated the successful rate in identifying an
event with the low uncertainty rate. Users were able
to successfully identify an event if they were certain
with the stories or information received. Figure 2
shows the number of successful event detected for
high and low stories across topics. There was no
statistical significance difference in the number of
successful event detected (Mann-Whitney Test,
p>0.05). However there was a statistical significance
difference in the number of successful event detected
(Mann-Whitney Test, p<0.05) in conjunction with the
low number of stories across topics. Users were able
to successfully detect an event even they were
provided with low number of stories. This occurred
0
50
100
150
200
250
300
P1 P2 P3 P4 P5 S1 S2 S3 S4 S5
Person Location Organisation
Date Money Percent
0
20
40
60
P1 P2 P3 P4 P5 S1 S2 S3 S4 S5
Low High
Exploring the Role of Named Entities for Uncertainty Recognition in Event Detection
339
in topics P3 (Gingrich Resigns), P5 (Clinton's Gaza
Trip) and S4 (Yankees vs. Padres in World Series).
One of the possibility was the high distribution of
named entities for these topics as shown in Table 6.
Topic P3 has the highest Location NEs (20.2%).
Topic P5 has the highest Person (19.4%), Money
(14.3%) and Percent (13.0%) NEs. Meanwhile topic
S4 has the highest Organisation (18.1%) and Date
(17%) NEs.
Table 6: Named entities distribution.
Topic Person Location Organisation Date Money Percent
P1 8.3 11.8 4.2 7.3 6.5 6.1
P2 12.7 7.4 9.4 10.6 13.5 9.6
P3 9.3 20.2 12.6 12.2 11.4 8.8
P4 5.6 8.4 12.1 11.0 14.1 12.9
P5 19.4 17.5 10.8 8.6 14.3 13.0
S1 6.0 7.7 6.3 14.9 8.0 10.4
S2 4.7 6.3 8.1 8.0 11.2 7.0
S3 8.2 6.7 9.8 6.7 4.3 8.7
S4 11.8 8.5 18.1 17.0 10.8 11.3
S5 14.1 5.5 8.6 3.7 5.9 12.2
Users managed to detect an event when they were
provided with high number of stories compared to
when they were provided with low number of stories.
However there was an exception if the low number of
stories have high number of NEs. This indicated that
named entities could reduce the uncertainty in event
detection although users were provided with low
number of stories.
4.2.2 Type of Information
Users issued an average of 22 keywords of NEs per
topic and an average of 4 keywords of NEs per event.
Meanwhile they also provided double the amount for
BOW; an average of 47 keywords of BOW per topic
and an average of 9 keywords of BOW per event. The
number of keywords that were labelled as ambiguous,
complex and conflict in event detection. It is
important to analyse the distribution of these
keywords to identify what type of information (NEs,
BOW) and in which condition (low, high) will
significantly contribute to uncertainty as shown in
Table 7.
There was no significant difference in type of
information across topics for complex and conflict
dimension (Mann-Whitney Test, p>0.05). Users
tends to provide almost the same average amount of
keywords between NEs (mean=14) and BOW
(mean=16) per topic.
However there was a statistical significance
difference for type of information (Mann-Whitney
Test, p<0.05) in conjunction with the low
information/stories across topics for ambiguous
dimension. User list out less NEs (mean=3.01,
sd=4.05) as an ambiguous information compared to
BOW (mean=9.42, sd=6.45) when they were
provided with low information/stories in topics P3
(Gingrich Resigns), P5 (Clinton's Gaza Trip) and S4
(Yankees vs. Padres in World Series). One of the
reason probably the high number of NEs occurred in
these topics has help user in event detection task.
Table 7: Type of information distribution across different
settings.
Frequency (%)
Topic
Amount
of info./
stories
Ambiguous Complex Conflict
P1
Low 34.7 (65.3) 54.3 (45.7) 44.1 (55.9)
High 45.7 (54.3) 50.2 (49.8) 48.2 (51.8)
P2
Low 28.1 (71.9) 44.7 (55.3) 54.4 (45.6)
High 42.4 (57.6) 54.1 (45.9) 46.7 (53.3)
P3
Low 15.6 (84.4) 47.9 (52.1) 55.8 (44.2)
High 51.2 (48.8) 50.5 (49.5) 55.2 (44.8)
P4
Low 23.7 (76.3) 47.8 (52.2) 44.9 (55.1)
High 47.1 (52.9) 45.5 (54.5) 54.7 (45.3)
P5
Low 12.6 (87.4) 54.5 (45.5) 47.7 (52.3)
High 40.5 (59.5) 47.4 (52.3) 50.7 (49.3)
S1
Low 29.6 (70.4) 54.8 (45.2) 47.5 (52.5)
High 50.3 (49.7) 47.8 (52.2) 45.9 (54.1)
S2
Low 31.7 (68.3) 44.2 (55.8) 47.2 (52.8)
High 53.3 (46.7) 55.3 (44.7) 44.4 (55.6)
S3
Low 25.5 (74.5) 49.9 (50.1) 54.2 (45.8)
High 48.9 (51.1) 47.1 (52.9) 45.8 (54.2)
S4
Low 19.8 (80.2) 55.1 (44.9) 47.6 (52.4)
High 52.1 (47.9) 45.6 (54.4) 50.4 (49.6)
S5
Low 33.5 (66.5) 54.6 (45.4) 54.8 (45.2)
High 46.2 (53.8) 50.0 (50.0) 47.2 (52.8)
* Figure in bracket referring to the frequency of BOW (%)
This indicates that user perceived BOW as the
ambiguous information when they were given with
low number of stories to detect an event.
4.2.3 Post Evaluation Interview
During the post-evaluation interview, 85% of the
users agreed that the BOW were more descriptive
thus making the event detection task difficult.
Meanwhile 92% of the users agreed that named
entities has produced interesting and specific
information which has helped them to become focus
in identifying an event even they were provided with
low number of stories.
98% of the users agreed that ambiguous, complex
and conflicting information has nothing to do with
their understanding of the meaning of a term. 95% of
the user claimed that concentrating on the named
entities appeared in a stories has help them to
successfully perform the task.
RDBPM 2015 - Special Session on Research and Development on Business Process Management
340
5 CONCLUSIONS
User are able to detect an event even when they were
provided with low information or stories. Low
percentage of NEs was labelled as ‘ambiguous’ by
user during the event detection task. Thus NEs reduce
ambiguity and uncertainty in event detection,
compared to bag of words which is more descriptive.
This is one of the justification that NEs increased the
user confidence in understanding the flow of stories
by providing user with high quality forms of
information. Associating NEs occurred in an event
could be one of user strategy to increase their
understanding of a topic. Therefore another future
direction lies is by analysing named entity recognition
and linking to reduce uncertainty.
REFERENCES
TDT Corpus. https://catalog.ldc.upenn.edu/LDC98T25.
TDT Annotation Manual. https://catalog.ldc.upenn.edu/
docs/LDC2006T19/TDT2004V1.2.pdf.
Mohd, M and Mabrook, O., 2014. Investigating the
Combination of Bag of Words and Named Entities
Approach in Tracking and Detection Tasks among
Journalists. Journal of Information Science Theory and
Practice. 2(4), pp 31-38.
Hurley, R. J. 2011. Uncertain about cancer? so is online
news. Communication Currents, 6 (5).
http://www.natcom.org/CommCurrentsArticle.aspx?id
=1703.
Chowdhury, S., Gibb, F., & Landoni, M. 2011. Uncertainty
in information seeking and retrieval: A study in an
academic environment. Inf. Process. Manage. 47, 2
(March 2011), pp. 157-175. DOI=http://dx.doi.org/
10.1016/j.ipm.2010.09.006.
Ingwersen, P. 1992. Information Retrieval Interaction,
Taylor Graham, London.
Goodman, N. D., Vikash, K., Mansinghka, D. R., Bonawitz,
K., and Tenenbaum, J.B. 2008. Church: A language for
generative models. In Proceedings of the Twenty-
Fourth Conference Annual Conference on Uncertainty
in Artificial Intelligence, Corvallis, Oregon, pp. 220-
229.
Topka, L. V., 2013. Situation of uncertainty: pragmatic,
semantic, and syntactic aspects of investigation.
European Scientific Journal. 9 (Sept. 2013), pp. 60-69.
Rubin, V. L., Liddy, E. D., and Kando, N. 2006. Certainty
identification in texts: categorization model and manual
tagging results. Springer, Dordrecht, The Netherlands,
vol. 20, pp. 61-76.
Goujon, B. 2009. Uncertainty detection for information
extraction. In Proceedings of the International
Conference RANLP 2009, Borovets, Bulgaria, pp. 118-
122.
Mishel, M. H. 1988. Uncertainty in illness. Image J Nurs
Sch, vol. 20, pp. 225–232.
Babrow, A. S., Kasch, C. R., and Ford, L. A. 1998. The
many meanings of uncertainty in illness: towards a
systematic accounting. Heatlh Communication, vol.10,
pp. 1-23.
Chen H. and Ku L. W. 2002. An NLP and IR approach to
topic detection. In: Allan J (ed.) Topic detection and
tracking: Event-based information organization.
Norwell, MA: Kluwer Academic Publishers, pp. 243–
264.
Cunningham, H., Maynard, D., Bontcheva, K., and Tablan,
V. 2002. GATE: A framework and graphical
development environment for robust NLP tools and
applications. In Proceedings of the 40th Anniversary
Meeting of the Association for Computational
Linguistics, July 2002, Philadelphia, PA, pp. 168–175.
Sparck-Jones, K., and Willet, P. 1997. Readings in
Information Retrieval. San Francisco, CA: Morgan
Kaufmann.
Exploring the Role of Named Entities for Uncertainty Recognition in Event Detection
341