Context Sensitivity in Knowledge Rich Systems Part III

Context Sensitivity in Knowledge Rich Systems Part III: Case Studies Igor Mozetič Jozef Stefan Institute, Slovenia November 2006 Tutorial at ISWC 2006, Athens, Georgia 1

Part III: Overview n n n Contextualized ontology Ontology mapping Multilingual context Ontology learning in context News analysis, temporal models as context November 2006 Tutorial at ISWC 2006, Athens, Georgia 2

Contextualized Ontology A Onto 1 What is A in the context of Onto 1 ? November 2006 Tutorial at ISWC 2006, Athens, Georgia 3

Contextualized Ontology fish Euro. Voc Asfa What is “fish” in the context of legal EU terminology ? November 2006 Tutorial at ISWC 2006, Athens, Georgia 4

Contextualized Ontology fish Food. Onto Asfa What is “fish” in the context of cooking ? November 2006 Tutorial at ISWC 2006, Athens, Georgia 5

Food. Ontology (W 3 C, OWL-guide) Fish as food Seafood Fish Bland. Fish November 2006 Shellfish Non. Bland. Fish Tutorial at ISWC 2006, Athens, Georgia 6

Euro. Voc Fish as a resource Fishery resources aquatic plant mollusc crustacean Fish plankton fish farming fish oil fish product fish desease sea fish freswater fish November 2006 Tutorial at ISWC 2006, Athens, Georgia 7

Ontology Grounding broader term fish related term narrower term doc 1: … fish… doc 2: … … (fish) doc 3: November 2006 Tutorial at ISWC 2006, Athens, Georgia 8

Ontology Grounding n Artificial Intelligence: Symbol grounding q q n symbols are grounded in perceptions [Harnad: The Symbol Grounding Problem, Physica D 42, 1990] Semantic Web: Ontology “grounding” q q ontology concepts are grounded in data (documents, web pages, database, …) [Jakulin & Mladenic: Ontology Grounding, Si. KDD, 2005] November 2006 Tutorial at ISWC 2006, Athens, Georgia Proc. 9

Part III: Overview n n n Contextualized ontology Ontology mapping Multilingual context Ontology learning in context News analysis, temporal models as context November 2006 Tutorial at ISWC 2006, Athens, Georgia 10

FAO case study: ASFA abstracts n n ASFA thesaurus Contents: q biblio info q abstracts, indexed by thesaurus terms Scope: q marine sc, tech, management, economy, social Numbers: q 1 mio documents q from 5000 publications (several external partners) November 2006 Tutorial at ISWC 2006, Athens, Georgia 11

FAO: ASFA thesaurus n n 6500 descriptors (permitted) 3500 non-descriptors (forbidden) relations: q BT broader term q NT narrower term q RT related term q use (instead), used_for, scope_notes, … English only November 2006 Tutorial at ISWC 2006, Athens, Georgia 12

FAO: ASFA abstract November 2006 Tutorial at ISWC 2006, Athens, Georgia 13

FAO: data and problems November 2006 Tutorial at ISWC 2006, Athens, Georgia 14

FAO: ASFA – Euro. Voc mapping Euro. Voc B A C Asfa map( A(Asfa), B(Euro. Voc), 0. 8 ) map( A(Asfa), C(Euro. Voc), 0. 65 ) November 2006 Tutorial at ISWC 2006, Athens, Georgia 15

Ontology Mapping n Find mappings between features of grounded concepts Onto 1 November 2006 A B features: parse tree, Bo. W, … grounding: doc 1, doc 2, … grounding: doc 3, doc 4, … Tutorial at ISWC 2006, Athens, Georgia Onto 2 16

Part III: Overview n n n Contextualized ontology Ontology mapping Multilingual context Ontology learning in context News analysis, temporal models as context November 2006 Tutorial at ISWC 2006, Athens, Georgia 17

Mapping Multilingual Ontologies n Given aligned multilingual docs, find mappings fish features grounding: English docs November 2006 pescado features grounding: Spanish docs Tutorial at ISWC 2006, Athens, Georgia languageindependent representation aligned multilingual documents 18

KCCA (Kernel Canonical Correlation Analysis) KCCA learns a semantic representation of the text [slide adapted from Fortuna & Shawe-Taylor 05, Vinokourov etal. 02] : n Input: set of paired documents (for each document there is a version in each language) n Output: set of mappings from native language space into “language independent space” – subspace with semantic dimensions November 2006 KCCA loss, income, company, quarter verlust, einkommen, firma, viertel Semantic dimensions wage, payment, negotiati-ons, union zahlung, volle, gewerkschaft, verhandlungsrunde Tutorial at ISWC 2006, Athens, Georgia 19

References n n n Fortuna, Cristianini, Shawe-Taylor: A Kernel Canonical Correlation Analysis For Learning The Semantics Of Text, in Kernel methods in bioengineering, communications and image processing, 2006. Fortuna, Shawe-Taylor: The use of machine translation tools for cross-lingual text mining Learning With Multiple Views, Proc. Workshop at 22 nd ICML, 2005. Vinokourov, Shawe-Taylor, Cristianini: Inferring a semantic representation of text via cross-language correlation analysis, Advances of Neural Information Processing Systems 15, 2002. November 2006 Tutorial at ISWC 2006, Athens, Georgia 20

Part III: Overview n n n Contextualized ontology Ontology mapping Multilingual context Ontology learning in context News analysis, temporal models as context November 2006 Tutorial at ISWC 2006, Athens, Georgia 21

Ontology Learning (in Context) n Given: Corpus of documents n Construct (semi-automatically): An ontology (concepts and relations) n Use: Contexts (other ontologies) to suggest names of concepts (and relations) November 2006 Tutorial at ISWC 2006, Athens, Georgia 22

Ontology Learning (in Context) Asfa A B ? Yahoo! Dmoz docs 3 docs 1 November 2006 Tutorial at ISWC 2006, Athens, Georgia 23

References Onto. Gen: n B. Fortuna, M. Grobelnik, D. Mladenic: System for Semiautomatic Ontology construction, Demo at ESWC 2006. n B. Fortuna, M. Grobelnik, D. Mladenic: Background Knowledge for Ontology Construction, Poster at WWW 2006 and at Workshop on Context & Ontologies, ECML 2006. November 2006 Tutorial at ISWC 2006, Athens, Georgia 24

Part III: Overview n n n Contextualized ontology Ontology mapping Multilingual context Ontology learning in context News analysis, temporal models as context November 2006 Tutorial at ISWC 2006, Athens, Georgia 25

News analysis n n Given: stream of news Tasks: q q q n Find relations between news Interpret unrelated news Predict future events Use: Models to provide the interpretation context November 2006 Tutorial at ISWC 2006, Athens, Georgia 26

News analysis News stream: earthquake waves tsunami related? explain? November 2006 the same? Tutorial at ISWC 2006, Athens, Georgia 27

News analysis: Temporal model = Context News stream: earthquake waves temporal model, spatial model, economic model, … November 2006 tsunami provides context for subsequent events Tutorial at ISWC 2006, Athens, Georgia 28

News analysis: Example n n Day 1: “Giant waves hit the shore early today. ” Day 2: “An ocean floor earthquake was detected yesterday. ” Levels of (semantic) similarity: 1. lexical (keywords) 2. lexicographic (taxonomies) 3. model-based (models of word referents) November 2006 Tutorial at ISWC 2006, Athens, Georgia 29

News analysis: Temporal model Interval algebra [Allen 83]: n Temporal relations (before, during, after, …) n Events = time intervals November 2006 Tutorial at ISWC 2006, Athens, Georgia 30

News analysis: Defeasible hypothesis n News consistent with the model: => Tsunami is a possible semantic link between the two events. n Counterexample: Waves before Earthquke => Inconsistent with the tsunami model November 2006 Tutorial at ISWC 2006, Athens, Georgia 31

References n n n Allen: Maintaining knowledge about temporal intervals, CACM 26, 1983. Molovan, Clark, Harabagiu: Temporal context representation and reasoning, Proc. IJCAI, 2005. Mozetic, Bojadziev: Reasoning with temporal context in news analysis, ECAI workshop on Context & Ontologies, 2006. November 2006 Tutorial at ISWC 2006, Athens, Georgia 32