A semantic contextdriven approach

1

A semantic context-driven approach for
georeference techniques Dario Saracino -
Exprivia S.p.A. Marco Cammisa Exprivia
S.p.A. Roberto Basili University of Rome Tor
Vergata Lucio Colaiacomo - EUSC
ESA-Eusc Conference Frascati, 05/03/2008
2
Welcome
3
Outline

• Introduction
• LSA based named entity recognition
• Experiments
• Results
• Conclusions

4
Motivation

• EUSC EKP Project an application to support the
  EUSCs image analysts
• GeoReference is important for the military
  intelligence framework, it allows to intercept
  relevant location in a text
• Attach to a location geographical attributes as
  latitude and longitude (NGA ex NIMA)
• GeoReference can be approached as a Named Entity
  Recognition task
• LSA as a method and theory for the Named Entity
  Recognition

5
Introduction

• Results are part of the application EUSC EKP
  Project
• Web spidering
• Classification
• Natural language query
• Geographical query
• Clustering of Web Pages
• Geo-Reference extraction
• Results are related to a subset of the total
  documents downloaded 1000 of 5500
• www.globalsecurity.org
• www.defenselink.mil
• www.intelcenter.com
• www.military.com

6
LSA Based Named Entity Recognition

• LSA is the acronym for Latent Semantic Analysis
• based on SVD (Singular Value Decomposition),
  factorization algorithm in linear algebra
  (applied in signal processing and statistics)
• extracts latent information observing the
  co-occurrence of phenomena

7
How LSA scales

• S(k,k) reduced to S(3,3), now the original
  matrix A is just approximated
• Exists a function F that allows to compare OBSERV
  and PHENOMENA

8
How the original is approximated

• Example of LSA applied on the wavelength of the
  image
• Increasing K, the original image is more
  approximated
• LSA can be considered as a loss informative
  algorithm
• The less informative information (including
  noises) are loss

9
Application to Named Entity Recognition

• LSA and word order Simon Dennis, "Introducing
  word order in an LSA framework 2007
• Focus on the word order of the contexts of the
  targeting words
• Example Unanimously adopting resolution 1747
  (2007), submitted by France, Germany and the
  United Kingdom, the Council affirmed its decision
  that Iran should, without further delay,
• Left context of France resolution17472007s
  ubmittedby
• Right context of France GermanyandtheUnite
  d Kingdomthe

10
Experiments

• 5500 web pages downloaded
• 4 classes of Named Entities location, person,
  community, military armament
• 1650 contexts validated
• Split train/test 20, 30, 40

11
Results

• We measured
• Precision the number of items correctly labeled
  divided by the total number of elements labeled
• Recall the number of items correctly labeled
  divided the number of elements that actually
  belong to the class
• F1-Measure weighted harmonic mean of precision
  and recall (F12PR/(PR))
• Baseline, 72.8 (the most likely category,
  location)

12
Conclusion

• GeoReferencing has been approached as a named
  entity recognition task
• A LSA based named entity recognizer has been
  evaluated on a real case of web pages retrieved
  from the web
• Performance in terms of Precision, Recall and
  F1-Measure
• Best result of 93 of F1-measure for the 40
  split train/test (baseline 72.8)
• Future works
• Evaluation of alternative approaches HMM,
  Maximum Entropy, hybrid,
• Increase the number of classes
• Wider evaluation