Discovering Semantic Sibling Groups from Web Documents with XTREEMSG

1
Discovering Semantic Sibling Groupsfrom Web
Documents with XTREEM-SG

• Marko Brunzel, Myra Spiliopoulou
• marko.brunzel_at_dfki.uni-kl.de
• Otto-von-Guericke-University Magdeburg and
• German Research Centre for Artificial
  Intelligence (DFKI GmbH)
• EKAW 2006, 05.10.2006

2
Content

• Introduction
• Group-By-Path
• XTREEM-SG Procedure
• Experimental Results
• Conclusion and Outlook

3
Context

• Semantic Web
• Ontologies
• Knowledge Acquisition Bottleneck
• Ontology Learning

4
Semi-Structured Data Source

• Structure
• Dictionaries, glossaries, database schemas
• Rarely available
• Unstructured Text
• 50 methods
• Shared conceptualization in written text?
• Semi-Structured Text
• Web Documents (X)HTML

5
WWW as Data Source

• Domain Document Collection
• As Input ? manually crafted ? laborious
• Subjective - limited confidential
• WWW Documents
• As part of the mining process
• Shared - all topics - freely available
• Added value of mark-up

6
Semi-Structured Web Content

• HTML Text Mark-up
• XHTML
• Enforces Tree Structure
• HTML to XHTML Conversion
• Mark-Up is used for
• Tables / Lists / Headings / (Named) Links /
  Highlightings

? XTREEM - Xhtml TREE Mining
7
Hints on Semantic Relatedness - Examples
Headings- distributed over paragraphs
lth2gtWordnetlt/h2gtltpgtWas developed
lt/pgtlth2gtGermanetlt/h2gtltpgtAnalogous lt/pgt
Highlighted Keywords - with regular text between
ltpgt there are different important standards
for building the ltstronggtSemantic Weblt/stronggt.
is ltstronggtRDFlt/stronggt. ltstronggtRDFS lt/stronggt
adds whereas ltstronggtOWL lt/stronggt is lt/pgt
8
Content

• Introduction
• Group-By-Path
• XTREEM-SG Procedure
• Experimental Results
• Conclusion and Outlook

9
XHTML-Tree
lthtmlgt ltheadgt ltheadgt lt/headgt
ltbodygt lth1gtLexical
Resourceslt/h1gt ltpgtlt/pgt
lth2gtWordnetlt/h2gt ltpgtWas developed
lt/pgt lth2gtGermanetlt/h2gt
ltpgtAnalogous to Wordnet for the English lt/pgt
ltbodygt lt/bodygt lt/htmlgt
10
XHTML-Tree Paths
lthtmlgt ltheadgt ltheadgt lt/headgt
ltbodygt lth1gtLexical Resources
lt/h1gt ltpgtlt/pgt
lth2gtWordnetlt/h2gt ltpgtWas developed
lt/pgt lth2gtGermanetlt/h2gt
ltpgtAnalogous to Wordnet for the English lt/pgt
ltbodygt lt/bodygt lt/htmlgt
lthtmlgt lthtmlgtltheadgt lthtmlgtltheadgt lthtmlgtlt/headgt lth
tmlgtltbodygt lthtmlgtltbodygtlth1gtLexical
Resourceslt/h1gt lthtmlgtltbodygtltpgtlt/pgt lthtmlgtltbodygtlth
2gtWordnetlt/h2gt lthtmlgtltbodygtltpgtWas developed
lt/pgt lthtmlgtltbodygtlth2gtGermanetlt/h2gt lthtmlgtltbodygtltp
gtAnalogous to Wordnet for the English
lt/pgt lthtmlgtltbodygt lthtmlgtlt/bodygt lt/htmlgt
11
XHTML-Tree Paths Text Elements
lthtmlgt lthtmlgtltheadgt lthtmlgtltheadgt lthtmlgtlt/headgt lth
tmlgtltbodygt lthtmlgtltbodygtlth1gtLexical
Resourceslt/h1gt lthtmlgtltbodygtltpgtlt/pgt lthtmlgtltbodygtlth
2gtWordnetlt/h2gt lthtmlgtltbodygtltpgtWas developed
lt/pgt lthtmlgtltbodygtlth2gtGermanetlt/h2gt lthtmlgtltbodygtltp
gtAnalogous to Wordnet for the English
lt/pgt lthtmlgtltbodygt lthtmlgtlt/bodygt lt/htmlgt
12
XHTML-Tree Paths Text Elements - Filtered
lthtmlgt lthtmlgtltheadgt lthtmlgtltheadgt lthtmlgtlt/headgt lth
tmlgtltbodygt lthtmlgtltbodygtlth1gtLexical
Resourceslt/h1gt lthtmlgtltbodygtltpgtlt/pgt lthtmlgtltbodygtlth
2gtWordnetlt/h2gt lthtmlgtltbodygtltpgtWas developed
lt/pgt lthtmlgtltbodygtlth2gtGermanetlt/h2gt lthtmlgtltbodygtltp
gtAnalogous to Wordnet for the English
lt/pgt lthtmlgtltbodygt lthtmlgtlt/bodygt lt/htmlgt
13
Group-By-Path
lthtmlgt lthtmlgtltheadgt lthtmlgtltheadgt lthtmlgtlt/headgt lth
tmlgtltbodygt lthtmlgtltbodygtlth1gtLexical
Resourceslt/h1gt lthtmlgtltbodygtltpgtlt/pgt lthtmlgtltbodygtlth
2gtWordnetlt/h2gt lthtmlgtltbodygtltpgtWas developed
lt/pgt lthtmlgtltbodygtlth2gtGermanetlt/h2gt lthtmlgtltbodygtltp
gtAnalogous to Wordnet for the English
lt/pgt lthtmlgtltbodygt lthtmlgtlt/bodygt lt/htmlgt
14
Content

• Introduction
• Group-By-Path
• XTREEM-SG Procedure
• Experimental Results
• Conclusion and Outlook

15
Sibling Sets

• Sibling Text-Spans - Item Sets
• Wordnet, Germanet
• Errors
• Large Amounts of Web Documents ? Redundancy ?
  Processing ? reduce sporadic occurring Siblings

16
XTREEM-SG Data Flow Diagram
17
Resulting Clusters
Publication Parts
TopicMap
topic association scope role associations subject
topics resource occurrence topic_map
abstract keywords title references acknowledgement
s speaker introduction contact conclusion authors
thesaurus taxonomy ontology metadata controlled_vo
cabulary faceted_classification semantic_web topic
_maps classification rdf
Metadata
ian_horrocks dieter_fensel stefan_decker steffen_s
taab frank_van_harmelen peter_f__patel_schneider d
eborah_l__mcguinness raphael_volz brian_mcbride se
an_bechhofer
xml rdf html owl daml_oil semantic_web rdf_schema
xml_schema http w3c
class property domain range subclassof mincardinal
ity cardinality disjointwith string list
wordnet cyc opencyc sensus sumo semanticweb_org on
tosaurus umls daml_oil shoe
Ontologies
Authors
Webservice
RDF/OWL
18
Evaluation

• Automatic vs. Manual
• Gold standard ontologies
• F-Measure on average sibling overlap
• Set of sets / set of sets

19
Variations on the Preprocessing method

• Group-By-Path (GBP)
• traditional Bag-Of-Words (BOW) vector space model
  
• solely usage of Mark-Up (MU)

20
Evaluation Results I
21

• FMASO
• More intuitive (since often seen before)
• precision and recall

22
Evaluation Results PR
23
Variations on the Web Document Collection
24
Evaluation Results II
25
Variations on the required support

• threshold on the required support
• of terms in the Web Document Collection
• weakly supported terms are more and more ignored

26
Evaluation Results III
27
Findings

• Improvement of FMASO from 14.18 to 21.47
• Cluster Characteristics
• Conventional relations between terms in a cluster
  are various
• XTREEM-SG Sibling Relationship
• Explanation
• Authors group texts at the same level into item
  lists, headlines etc, - usually motivated by the
  intention to present sibling concepts in an
  intuitive way

28
Content

• Introduction
• Group-By-Path
• XTREEM-SG Procedure
• Experimental Results
• Conclusion and Outlook

29
Conclusion and Outlook

• Discovery of
• Siblings (co-hyponyms, co-meronyms, )
• Singleword multiword term expressions
• Language and domain independent
• WebScale
• Large scale application / integration
• number of clusters to be generated/inspected
• Corresponding super-concept

30

• Thank you for your attention!
• Questions?