Arabic Event Detection and Tracking

1
Arabic Event Detection and Tracking

2
Outline

• Arabic Event Detection
• Detection System
• Experiments
• Arabic Event Tracking
• Data
• Approaches
• Experiments

3
Arabic Event Detection

• Goal
• Group together stories that discuss the same
  event.
• CMU Detection System
• GAC group average clustering algorithm
• INCR a single-pass incremental clustering
  algorithm

4
Arabic Detection Experiments

• Data
• TDT3
• Arabic news stories from 1998/10 1998/12
• TDT4
• Arabic news stories from 2000/10 2001/01
• Detection Tasks
• Arabic, English Translations
• Arabic, Native Orthography

5
Results Arabic Event Detection

• TDT 2002 dry-run evaluation result

Condition Cost
TEArb,Eng 0.1770
TEArb,Nat 0.1732
6
Results Arabic Event Detection

• TDT 2002 formal run evaluation result

Condition Cost
TEArb,Eng 0.2450
TEArb,Nat 0.2437
7
Arabic Event Tracking

• Data
• Approaches
• Evaluation
• DET curve
• Experiments

8
Data -- Arabic Event Tracking

• Data
• TDT3
• Task
• Training set English
• Test set Arabic
• Events
• 23 events

9
Approaches Arabic Event Tracking

• Translating test set
• IBM MT system (released by NIST)
• Translating training set
• English-Arabic bilingual dictionary (statMT) we
  automatically extracted from UN English-Arabic
  parallel corpus

10
Experimental Results Arabic Event Tracking

• Translating test set

Condition Cost
TDT2002 dryrun best result for multilingual task 0.1307
CMU Arabic task (after adaptation) 0.1364
11
Experiments Results Arabic Event Tracking

• Adaptation Effect

Condition Cost Reduction Ration
Without Adaptation 0.1429 --
With Adaptation 0.1364 4.5
12
Experimental Results Arabic Event Tracking

• Translating training set

Condition Cost Reduction Ratio
translating training set by DICT 0.1772 --
translating training set by DICT Adaptation 0.1633 7.8
13
Experiment Results Arabic Event Tracking