The CROSSMARC Platform for Web Information Retrieval and Extraction

1
The CROSSMARC Platform for Web Information
Retrieval and Extraction
Software Knowledge Engineering
Laboratory Institute of Informatics
Telecommunications NCSR Demokritos, Athens,
Greece
Vangelis Karkaletsis, Constantine D. Spyropoulos

NEMIS, Athens, 25 October 2004
2
Contents

• Problem Description
• The CROSSMARC project
• The Platform
• Customization Infrastructure
• The customizable run-time system
• Application Building
• Benefits
• Conclusions

3
Problem Description I
4
Problem Description II

• Poor performance of general-purpose search
  engines, which depend on the results of generic
  Web crawlers.
• The behavior of the search engines must be
  adapted to the users requirements.

5
Problem Description III

• The traditional approach to web IE involves the
  creation of wrappers for specific web sites.
• Manual creation of wrappers presents many
  shortcomings due to
• the overhead in writing and maintaining them
• Automatic creation of wrappers (wrapper
  induction) presents also problems since wrapper
  re-training is necessary
• when changes occur or
• when pages from a similar Web site are to be
  analysed
• Wrappers are more appropriate for structured or
  semi-structured pages

6
Wrappers Definition

• Wrapper creation problem can be stated as follows
  (Laender et al. 2002)
• given a Web page W containing the information of
  interest determine a mapping M that populates a
  data repository R with this information,
• the mapping M must be able to extract data from
  any other page W similar to W, where similar
  means in most cases belonging in the same site
  and having the same layout.

7
CROSSMARC objective
Develop a platform for the collection of web
pages and the extraction of information from
them, that

• enables handling of structured, semi-structured
  or unstructured data,
• enables adaptation to new domains and languages,
• facilitates maintenance for an existing domain,
• ensures personalised access to the extracted
  information,
• provides strategies for effective site navigation

8
CROSSMARC consortium
Completion Date 31/08/2003 Contract No IST 2000
25366
9
CROSSMARC Platform I

• Composed of
• a customization infrastructure that supports
  configuration to new applications and languages
• a run-time system for web information retrieval
  and extraction which can be trained using the
  customization infrastructure

10
CROSSMARC Platform II Customization
infrastructure

• Ontology Management
• Corpus Formation
• Corpus Collection and Annotation

11
Customization infrastructure Ontology management

• Ontology editor
• Creation and maintenance of domain ontologies
• Lexicon editor
• Creation and maintenance of domain lexica
• NERC editor
• Creation of the NERC DTD
• Template editor
• Creation of the FE template
• Stereotypes editor
• Creation and maintenance of user stereotypes

12
Customization Infrastructure Corpus Formation

• Corpus Formation is based on an interactive
  process between the user and a simple machine
  learning based classifier.

13
Customization Infrastructure Corpus Collection
and Annotation

• The corpus collection methodology determines
• how different web pages characteristics must be
  taken into account and
• how they are to be represented in the corpora.
• According to this methodology, web pages in each
  domain are classified in categories

14
Categories of Web pages containing Offers
Descriptions (2nd Domain)
15
Customization Infrastructure Corpus Collection
and Annotation

• The corpus collection methodology follows
  standard annotation practices for information
  extraction
• The annotation task is based on guidelines that
  are issued for each new domain and on the use of
  an annotation tool

16
CROSSMARC Platform III architecture of the
trainable system
Ontology
17
CROSSMARC Platform IV The trainable system
- Web pages collection

• Focused Crawler identifies web sites that are of
  relevance to a particular domain
• Combines 3 distinct crawler types
• Filters the list of Web sites produced
• Site-specific spider navigates in a Web site to
  locate domain specific pages

18
Web Pages Collection Focused Crawler
19
Web Pages Collection Site-specific spider

• Site navigation traverses a Web site, collecting
  information from each page visited and forwarding
  it to the Page-Filtering and Link-Scoring
  modules
• Page-filtering is responsible for deciding
  whether a page is an interesting one and should
  be stored or not
• before storing a page, its language is identified
• the page is also converted to XHTML
• Link-scoring validates the links to be followed.
  Only links with a score above a certain threshold
  are followed.

20
Web Pages Collection Site-specific spider
21
CROSSMARC Platform V The trainable system -
Information Extraction
22
Information Extraction different approaches of
the monolingual IE systems

• English IE machine learning based NERC,
  heuristics demarcator, classification based FE,
  name matcher normaliser
• Greek IE machine learning based NERC, heuristics
  demarcator, STALKER-based FE, name matcher
  normaliser
• Italian IE rule based NERC, heuristics
  demarcator, name matcher normaliser,
  WHISK-based FE, normaliser
• French IE hybrid NERC, heuristics demarcator,
  name matcher normaliser, hybrid FE

23
Information Extraction an example, the Greek IE
24
CROSSMARC Platform VI The trainable system
- Data Storage Presentation

• Data storage component
• stores the extracted facts, from the XML file
  produced by IE, into domain-specific databases.
• Data presentation is implemented by the end User
  Interface, an internationalized web-based
  application
• Personalization is performed by the
  general-purpose NCSR personalization server
  (PServer)

25
Application Building I

• Development of three applications to extract
  information from
• laptops offers in e-retailers web sites (in four
  languages),
• job offers in IT companies web sites (in four
  languages),
• holidays packages in the sites of travel
  agencies (in two languages)

26
Application Building II

• Involves two main stages
• Creation of application specific resources using
  the customization infrastructure
• Training of the system components using the
  application specific resources, configuration of
  the system components

27
Application Building III

• Stage 1 Creation of application specific
  resources
• Creation of concepts, their relationships and
  attributes using the ontology editor
• Creation of their linguistic realizations using
  the lexicon editor
• Specification of the important named entity types
  using the NERC editor to form a common DTD
• Specification of the FE XML schema using the
  template editor
• Collection of the corpus for page filtering using
  the corpus formation tool
• Collection and annotation of the corpus for IE
  using the annotation tools
• Specification of user stereotypes using the
  stereotypes editor

28
Ontology editor
29
Lexicon editor
30
NERC editor
31
Shared DTDs
Domain 1 NE MANUF, MODEL, PROCESSOR,
SOFT_OS TIMEX TIME, DATE,
DURATION NUMEX LENGTH, WEIGHT,
SPEED, CAPACITY, RESOLUTION,
MONEY, PERCENT

Domain 2 NE MUNICIPALITY, REGION,
COUNTRY, ORGANIZATION, JOB_TITLE,
EDU_TITLE, LANGUAGE, S/W TIMEX DATE,
DURATION NUMEX MONEY TERM SCHEDULE, ORG_UNIT
32
Template editor
33
FE schemas for 1st and 2nd domain
34
Stereotypes editor
35
Application Building IV

• Stage 2 Training of the system components,
  configuration
• Training of the Crawler
• Training of the page filtering and link scoring
  modules using the collected corpus
• Training of each monolingual IE sub-system using
  the collected corpus
• Customization of the UI exploiting the ontology,
  the lexicons and the stereotypes definitions
• Configuration of the system components

36
Benefits I

• CROSSMARC is not just another web extraction
  system
• CROSSMARC is a platform that provides a
  customization infrastructure and a trainable
  system
• To cope with the shortcomings of existing
  wrappers, CROSSMARC combines
• Wrapper Induction techniques to exploit the
  formatting features of the web pages,
• NLP techniques to exploit linguistic features of
  the web pages,
• Machine Learning techniques to facilitate
  customization to new applications
• enabling the process of application specific web
  pages in different sites and in different
  languages (multilingual, site-independent).
• CROSSMARC also employs ontology engineering
  techniques to coordinate the creation and
  maintenance of application and language specific
  resources

37
Categorization of tools for Web IE (Laender et
al. 2002)
38
Benefits II

• After evaluation performed during applications
  building, we can conclude that
• Crawler increased effort must be put into the
  initial stage of forming hypotheses about what
  would be good directory and query start points.
• Spider we are able to identify with a fairly
  high degree of confidence, when a Web page is an
  interesting one.
• Information Extraction satisfactory performance
  especially for offer descriptions extracted from
  simpler web pages. In addition, the existing
  systems can be tuned further in order to achieve
  better performance.

39
Crawler Evaluation

• more than one experimentation cycle may be needed
  depending on the domain and language

40
Spider Evaluation
41
Information extraction Evaluation

• Results comparable to MUC
• Full IE is a complex task which becomes even more
  complicated by the visual nature of web pages

42
Concluding remarks

• CROSSMARC is an operational platform for
  site-independent and multilingual information
  retrieval and extraction from web pages
• The run-time system is accessible from the
  CROSSMARC site.
• Access to various project resources and corpora
  will be provided for research purposes.
• Currently, more advanced components are tested
  and will be integrated in the platform.

43
Useful Links

• CROSSMARC site
• http//www.iit.demokritos.gr/skel/crossmarc
• Ellogon
• http//www.ellogon.org