Cláudio Baptista, UFCG A Model for Geographic Knowledge Extraction on Web Documents Cláudio E. C. Campelo and Cláudio de Souza.

Cláudio Baptista, UFCG http://lsi.dsc.ufcg.edu.br

A Model for Geographic Knowledge Extraction

on Web DocumentsCláudio E. C. Campelo andCláudio de Souza Baptista

University of Campina GrandeComputer Science Department

Information Systems Laboratoryhttp://www.lsi.dsc.ufcg.edu.br

SECOGIS – ER 2009Gramado – RS- Brazil, 13th November 2009


2

Agenda

Introduction Main Challenges Detection of Geographic References The Geographic Scope GeoSEn Prototype

Architecture GUI

Experiments Conclusion and Future Work


Introduction

Web: need for searching using the geographic context;

Traditional search engines: search based on keywords only;

Example: A Web document: “...With the arrival of the industry in

Gramado, one thousand of new jobs for Java programmers will be created...”;

User query: “Java programmer jobs Brazil”; The mentioned document will not be retrieved in

the previous query!


Introduction

What is the Geographic Context of Web documents? The place where the information was created? The places mentioned in the document content? Where are people who are most interested in a particular

information? etc…

Several documents have this context: Research in Portugal in which only occurrence of names

of Portuguese cities was considered (308 in total): Total of about 4 millions pages analyzed. Occurrence of 2.2 references per document; 4% of the queries submitted had a reference to one of those

cities.


Main Challenges

Detection of geographic references in the documents;

Modeling of geographic scope of documents;

Relevance ranking according to geographic context;

Need for efficient index techniques which cope with both textual and spatial dimensions

Development of user interfaces which provide usability to deal with both dimensions


Detection of Geographic References Aim: to identify document features which may be

mapped to a geographic place name; Challenge: elimination of ambiguities, ex:

Place with a name of a thing; (Ex. Gramado, Canela) Place with name of a Person (Ex. Garibaldi); Places with same names and same types: (Ex.

Cachoeirinha-Pe e Cachoeirinha-Rs); Places with same names and different types (ex. city

of Rio de Janeiro and state of Rio de Janeiro Places and gentilics with the same names (ex. city of

Paulista-Pe and paulista (who is born in São Paulo)


Detection of Geographic References Another example of ambiguity:

São Paulo as a State São Paulo as a City São Paulo as a football team São Paulo as the name of a hospital São Paulo as the Saint!


Detection of Geographic References Explored detected points: page content, page

title, URL; Types of detected places: all of the spatial

hierarchy: (from city to region); Types of detected references: place names,

postal code, telephone code area, gentilic.


Definitions

Confidence Rate (CR) represents the probability of a given reference be a valid place name.

Confidence Factor (CF) a measure associated to each analyzed feature during the detection of geographic reference.

CR

CF

1

N


Confidence Factor

CFST – analyzes the occurrence of special terms associated to geographic references; Examples of STs include: “in" (e.g. “in Gramado); "city"

(e.g. "city of São Paulo"); “ZIP” (e.g. “ZIP: 58109-000”); Storage of special terms:

Term; Type of geographic reference (zip code, telephone area

code, place name, etc,); Type of place (city, state, region); Minimum distance (DMIN);

Maximum distance (DMAX);

Maximum confidence grade (CMAX).


Confidence Factor CFTS – considers the probability of a term be a

geographic reference using a traditional search engine;


Confidence Factor

CFCROSS : analyzes the occurrence of cross references based on

topological relationships (inside, contains, etc);

CFFMT – evaluates the syntax used to describe the geographic references; Abbreviation of place names (R. de Janeiro, RJ); The use of uppercase in the place names; Telephone format ( 083)-999-3456; Postal code format 58.104-867


13

Modeling of the Geographic Scope A document may be associated to one or more

places; A geographic scope may have places that are

not mentioned directly in a document (geographic expansion)

Each place which is part of the scope has an associated relevance value;


14

Geographic Dispersion Rate

(a) (b)

Another factor used in the composition of the geographic relevance value;

Hypothesis: references dispersed may characterize regions that share common features (e.g. cultural, economic, social);


15

GeoSEn – an overview

Geographic Search Engine: Indexes a subset of the Brazilian Web; Deals with 6,291 places in Brazil, which are

organized in a five-levels hierarchy: from city to region.

Region: ex. South State: ex. Rio Grande do Sul MesoRegion: ex. Metropolitana de Porto Alegre MicroRegion: ex. Gramado-Canela Municipality: ex. Gramado


16

GeoSEn - Architecture



Query Example

Example of query using a user defined area of interestSELECT idFROM places plc1WHERE within(plc1.geometry, specified_geometry)AND NOT EXISTS ( SELECT id FROM places plc2 WHERE within(plc2.geometry, specified_geometry) AND within(plc1.geometry, plc2.geometry))


Experiments Experiments using 66,531 indexed

documents; 5 classes: .edu, .gov, blogs, tourism, arts; Detection of terms:

Documents from the Web manually analyzed; Documents with strong ambiguities created for the

test bed;


20

Conclusion

We have presented a heuristic based approach to implement a GIR system.

The techniques presented may be combined with others already known.

Precomputed relevance values may be used aiming to simplify the search process;


21

Future Work

Retrieval of georeferenced images and videos;

Recognition of other kinds of places; Integration of other data sources; Evaluation using large data set collections.


22

Thank you very much!

Questions?