slides - U

advertisement
Temporal Event Map
Construction For Event Search
Qing Li
Department of Computer Science
City University of Hong Kong
Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Algebra operations devised on TEM
 Conclusions and Future Work

Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Algebra operations devised on TEM
 Conclusions and Future Work

Introduction


Many news articles report events on the WWW
For an event, it may consist of several component events, i.e.,
episodes



There are relationship between component events
Some component events are more important than others
For example, “Toyota 2009-2010 vehicle recalls”

What people want from news articles?
Not a sole news article,
but the events reported by
some related news articles
Dependent relationships
between
component events
Which component events
play important roles
in the event evolution
or development

They are interested in the whole picture of an event
evolution or development along a time line



Read news is an important way to know what happen



includes the dependent relationships between component events
Event importance in the event evolution or development
Too many news
Too many topics
Time consuming job to read news to find what user wants





Understand what happens
What is important component events
What is the relations between component events
Temporal event map
Provide a convenient way to browse the event evolution
Introduction

Current search engine




Keyword search
List of web pages
Can not provide a map of an event
Necessary and useful to provide
temporal event map (TEM)
Previous Work

Qiaozhu Mei and Chengxiang Zhai. Discovering Evolutionary Theme Patterns from Text: An
Exploration of Temporal Text Mining. In Proceeding of the 11th ACM SIGKDD International
Conference on Knowledge Discovery in Data Mining, pp. 198-207, 2005.

Christopher C. Yang, Xiaodong Shi, and Chih-Ping Wei. Tracing the Event Evolution of
Terror Attacks from On-Line News. In Proceeding of the ISI 2006.

Jiangtao Qiu et al. Timeline Analysis of Web News Events. In Proceeding of the ADMA
2008.

Christopher C. Yang, Xiaodong Shi, and Chih-Ping Wei. Discovering Event Evolution Graphs
From News Corpora. IEEE Trans. Sys. Man Cyber. Part A 2009.

Jin, Peiquan et al. TISE: A Temporal Search Engine for Web Contents. Proceedings of the
2008 Second International Symposium on Intelligent Information Technology Application

A. Feng and J. Allan. Incident threading for news passages. In CIKM ’09: Proceeding of the
18th ACM conference on Information and knowledge management, pages 1307–1316, New
York, NY, USA, 2009. ACM.
Limitations of Previous Work




Only take time sequence and content similarity between two
component events into consideration
Using such two factors are not enough to identify dependent
relationships
Do not provide a way to measure event importance so as to
identify milestone events that are more interested by most users
Do not provide a convenient way to browse event evolution
Our Work



Formalize the problem of event search
Propose a framework to search events based on
users’ queries
Characteristics and contributions of our work

Characteristics and contributions of our work



Content dependent relationship - mutual information VS.
content similarity
Event reference relationship- some news articles of an event
may refer to (mention) other events
Adopt three kinds of event relationships which are temporal
relationship, content dependent relationship and event
reference relationship to identify a dependent relationship
between two events

Characteristics and contributions of our work




Define some algebra operations to assist user browsing TEM
The search results are organized by a temporal event map (TEM)
which is a whole picture about an event’s evolution or development
along a time line
Propose a method to measure event importance degrees so as to rank
events based on their importance degrees
Experiment results show that our method outperforms baselines in
discovering event dependent relationships and ranking events based
on event importance
Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Algebra operations devised on TEM
 Conclusions and Future Work

Event Modeling


Event – reported by documents
Document - A story talking about an event including the
happen time, places and content of the event

Timestamp
Places

Content

Event Modeling
Event Modeling

Related document set of an event - a set of documents talking
about the event

Life cycle

Begin time


End time


the earliest timestamp among all timestamps of related documents of
the event
latest timestamp among all timestamps of related documents of the
event
Place set
Event Modeling

Example: SARS epidemic




The life cycle of this event is from November 2002 to May
2006.
The places of the event includes China, Canada, Singapore and
so on.
There are many news from the world wide web which reported
such an event.
We can extract keywords from the set of documents to
describe the event such as SARS, flu-like, fever, treatment and
so on.
An Example of TEM
Our Work

Input


User can search events by time, places and interested content
Output

A temporal event map




Event evolution
Fuzzy relations of events
Important events
Some algebra operations
Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Algebra operations devised on TEM
 Conclusions and Future Work

Steps of Constructing Temporal Event Map
For Temporal Event Search






Identify related document set of target event
Event Discovering
Content Dependent Relationship Analysis
Event Reference Relationship Analysis
Event Ranking
Temporal Event Map Construction
Identify related document set of target event

Identify related document set of target event


The input information could be considered as the search requirements
of the user and corresponds to a target event which satisfy all the
requirements.
The related document set of the target event can be obtained by a
function.
Identify related document set of target event

Special cases of input


Partially input - part of (It, Ip, If)
We consider these input as three kinds of requirements and
only take the input requirements into consideration.
Rtf
Rtpf
1
2
3
4
5
1
2
4
5
Event discovering




For each target event a corresponding to an input I and its
related document set Ra, we can detect several sub-vents
from Ra.
We do not aim at event detection
Adopt the topic-model based method to detect the events
A function
Content Dependent Relationship Analysis

Event Representation – salient feature vector
Content Dependent Relationship Analysis



Previous works - content similarity
Some keywords in two events are dependent but not exactly matched
Calculating mutual information to measure the dependence between
features, and then use an aggregation of all mutual information between
features in events
Event Reference Relationship Analysis
Event Ranking


some component events are more important than others
ranking function to rank all the sub-events
0.82
0.97
E1
E5
E3
0.89
0.8
0.52
E2
0.66
E4
0.88
0.6
E6
Temporal Event Map Construction
Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Algebra Operations Devised on TEM
 Conclusions and Future Work

Algebra Operations Devised on TEM
Projection
Relation
 Projection
 Projection
 Projection
 Zoom In
 Zoom out

Based on Content Dependency
Based on Reference Relation
on Happening Places
on A Time Period
An example of the whole TEM about the event
“2011 Japan Earthquake”
An example of Projection Based on Content Dependency
Relation about the event “2011 Japan Earthquake”
An example of Projection Based on Reference Relation on
TEM about the event “2011 Japan Earthquake”
Projection on Happening Places on TEM about
the event “2011 Japan Earthquake”
Projection on Time Period on TEM about the
event “2011 Japan Earthquake”
Zoom In Operation
Sub-TEM of Zoom In “The earthquake and tsunami
caused a number of nuclear accidents.”
Outline
Introduction
 Problem Formulation
 Temporal Event Map Construction For
Event Search
 Evaluation
 Conclusions and Future Work

Conclusion and future work

Conclusion






Formulate the temporal event search problem
Propose a framework to search events according to users’ queries.
Define three kinds of relationships and use them to identify event dependent
relationships
The search results is represented by a temporal event map (TEM)
A method to measure event importance degree
Future work




Try to handle different kinds of input and discuss their scalability
Discover the important and burst periods of an event
To achieve personalization
News recommendation to let user know the event more clearly and completely
Thanks!
Download