©2004, Philippe Cudré-Mauroux Semantic Interoperability for Global Information Systems Microsoft Research Asia 08.20.04 Philippe Cudré-Mauroux Distributed

©2004, Philippe Cudré-Mauroux

Semantic Interoperability for Global Information Systems

Microsoft Research Asia 08.20.04

Philippe Cudré-Mauroux

Distributed Information Systems Laboratory (LSIR)Swiss Federal Institute of Technology, Lausanne (EPFL)


Outline

I. Classical Information Integration (overview)

– Global Schema– Multidatabase Language Approach– Federated Databases

II. Information Integration in the Large– Context: The Semantic Web– Shared ontologies– The Chatty Web

III. State of the Art in Ontology Alignment (overview)

IV. Semantic Integration in a Large-Scale Image Sharing Scenario


I. Classical Information Integration

• Goal: providing a uniform access to multiple heterogeneous information sources

• More than data exchange (e.g., ASCII, EDI, XML)

• Old problem, difficult, well-know (partial) solutions


Global Schema Integration

• Merge multiple databases into one global database

• Performed by human expert• Time consuming and error prone• Local autonomy lost• Static solution

Book(ISBN, Title, Price, Author)Author(Name, ISBN)

Livre(ISBN, Prix, Titre)Auteur(Prenom, Nom, ISBN)

Book(ISBN, Title)Author(Name, ISBN)

S1 S2


Multidatabase Language Approach

• No attempt at integrating schemas• Language (e.g., MSQL) used to integrate

information sources at run-time• Simple example:

• Not transparent• Heavy burden on (expert) users• Global queries subject to local changes

Use S1, S2Select TitreFrom S1.Book, S2.LivreWhere S1.Book.ISBN = S2.Livre.ISBN


Federated Databases

• Idea: Each information source exports a schema specifying shared relations

• Tight-coupling:– Global schema integration on all export schema (cf.

global schema integration)

• Loose-coupling:– Dynamic add / drop, e.g., by creating views (logical

relations)


GAV (Global as View)

• Global (mediated) schema is expressed as a view on local schemas

Book(ISBN, Title, Author)

[…][…]Book(ISBN, Title)Author(Name, ISBN)

Create VIEW Book AsSelect ISBN, Title, AuthorFrom S1.Book, S1.Author Where Book.ISBN = Author.ISBN

Mediated Schema

S1 S2 S3


LAV (Local as View)

• Local schemas are expressed as a view on global schema

Book(ISBN, Title, Author)

[…][…]Book(ISBN, Title)Author(Name, ISBN)

Create VIEW S1.Book AsSelect ISBN, TitleFrom Book

Mediated Schema

S1 S2 S3


LAV / GAV (cont.)

• Transparent access to heterogeneous databases in the federation

• Local autonomy is (usually) preserved• Query processing through query reformulation• Requires global agreement on the mediated

schema (tight semantic coupling)• Does not scale well


II. Information Integration in the Large

• Goal: providing a uniform access to many heterogeneous information sources

• Traditional approaches are inadequate– Lack of adaptability– Lack of transparency– Lack of scalability

• Hot research area


Some Applications

Agent Communication Web services integration Information retrieval from heterogeneous

databases Catalog matching P2P information sharing Personal information delivery Vertical information publishing


General Context: The Semantic Web

Unicode

XML + NS + xmlschema

RDF + rdfschema

Ontology vocabulary

Logic

Proof

Trust

URI

Dig

ital

Sig

nat

ure

Self-desc.doc.

Data

Data

Rules

• Providing machine-processable data to the Web


RDF/RDFS 2’ Overview

• RDF triple:

• RDF Schemas– Classes of resources– Classes of properties– Constraints on the subject (domain) or object (range)– Subclassing

• Extensible!

– Full-fledged ontological language: OWL

Subject ObjectProperty


Example: CreativeCommons

<rdf:RDF xmlns="http://web.resource.org/cc/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"><Work rdf:about="http://example.org/gnomophone.mp3"> <dc:title>Compilers in the Key of C</dc:title> <dc:description>A lovely classical work on compiling code.</dc:description> <dc:creator><Agent> <dc:title>Yo-Yo Dyne</dc:title> </Agent></dc:creator> <dc:rights><Agent> <dc:title>Gnomophone</dc:title> </Agent></dc:rights> <dc:date>1842</dc:date> <dc:format>audio/mpeg</dc:format> <dc:type rdf:resource="http://purl.org/dc/dcmitype/Sound" /> <dc:source rdf:resource="http://example.net/gnomovision.mov" /> <license rdf:resource="http://creativecommons.org/licenses/by-nc-nd/2.0/" /> <license rdf:resource="http://www.eff.org/IP/Open_licenses/eff_oal.html" /></Work>

<License rdf:about="http://creativecommons.org/licenses/by-nc-nd/2.0/"> <permits rdf:resource="http://web.resource.org/cc/Reproduction" /> <permits rdf:resource="http://web.resource.org/cc/Distribution" /> <requires rdf:resource="http://web.resource.org/cc/Notice" /> <requires rdf:resource="http://web.resource.org/cc/Attribution" /> <prohibits rdf:resource="http://web.resource.org/cc/CommercialUse" /></License></rdf:RDF>


Semantic Interoperability in The Semantic Web

• Common ontologies provide for shared context– Requires global agreement!

• Intractable standardization effort!• Back to stage 1…

• Two Plausible solutions:– Agreed-upon corpuses of basic concepts

• IEEE SUMO• Stanford TAP• …

– Local federation of ontologies fostering global interoperability• EPFL Chatty Web• U. Washington Piazza• …

• Complementary approaches


The Chatty Web

A lab in Trondheim

species

species

EMBLChange site at Cambridge

Swissprot siteat Geneva

A lab at MIT

organism

Query postedat EPFL

organism

organism

EMBLChange peersspecies, …

SwissProt peersauthors, titles, organism, …

other peersauthors, …

organism authors

organism species

species organism

• Local translations enabling global agreements

• Analyzing transitive closures of local mappings


On Translations Links (ontology mappings)


III. State of the Art on Ontology Alignment

• Problem: Given two ontologies which describe each a set of discrete entities, find the relationships holding between the entities

• Alignments can then be used to foster interoperability locally

• Difficult problem (fully automatic solutions?)

• Active area of research


Local Ontology Alignment Techniques

1. Terminological methods– string-based– language-based

• Intrinsic• Extrinsic• Multilingual

2. Structural methods– Internal– External

3. Others– Extensional– Semantic– User Feedback


1. Terminological Methods

• String-based: compare labels of entities– (sub-) String equality– Edit distances– Token-based distances (e.g., TF/IDF on substrings)

• Language-based– Intrinsic

• Terminological matching with morphological / syntactic analysis (allomorphies)

– Extrinsic• Use of external resources (e.g., WordNet synsets)

– Multilingual methods• Matches terms in different languages


2. Structural Methods

• Internal (constraint-based):– Data-based domain comparison– Multiplicities / Properties comparison– Similarity between collections

• External– Mereologic structures– Taxonomic structures– Relations bw similar entities


3. Other

• Extensional methods– Extension set of instances of a class

– Jaccard similarity:

– Similarity-based extension comparison

• Semantic Methods– Based on model-theoretic semantics– SAT problem (e.g., subsumption)

• User Feedback

)(

)(),(

BAP

BAPBA


A Handful of Systems

• APrompt (Stanford) [T,I,S,U]• Cupid (Microsoft research) [T,I,S]• Bibster (U. Karlsruhe) [T,I,S]• Glue (U. Washington) [E]• S-Match (U. Trento) [T,S,M]• …

Typically: a mix of techniques

[Terminological, Internal structure, external Structure, Extensional, seMantic, User]


IV. Semantic Integration in a Large-Scale Image Sharing Scenario

• Problem: retrieve a specific image from a large collection of shared images

• So far: most application mix CB and text analysis– CB image analysis provides a low-level objective

representation of an image• Good for comparing image features• Not so good w.r.t. end-users needs expressed in N.L.

– Surrounding text / filenames might sometimes be a high-level subjective view of the image

• Incomplete, out-of-context description• Good w.r.t. N.L. (cf. Google images)


Potential Opportunity

• Emerging applications make use of high-level, local and semi-contextualized image metadata– Structured metadata (Photoshop Album, XML)– Ontological metadata (RDF, Adobe XMP)– Type-based metadata (Microsoft WinFS)

• Paradigm shift from the old metadata standards (e.g., keywords, EXIF)– Extensible formats

• Personal conjecture: – Metadata will be prominent in a few years– Huge opportunity for image retrieval


Structured Metadata

• Ex.: Photoshop Album• Hierarchy of tags• Stored in a relational,

proprietary, local database

• Non-exportable


Ontological Metadata (1)

• Ex.: Extensible Metadata Platform (XMP)• Subset of RDF/S• Metadata might be embedded into the file• Supported by a wide range of Adobe applications

– Adobe® Acrobat®– Adobe FrameMaker® – Adobe GoLive®– Adobe Illustrator® – Adobe InCopy®– Adobe InDesign®– Adobe LiveMotion™– Adobe Photoshop®– Adobe Document Server – Adobe Graphics Server – Version Cue™


Ontological Metadata (2)

• Ex.: Photoshop XMP schema


Type-Based Metadata (1)

• New file-system for Longhorn (NTFS+++)• No more hierarchies (i.e., folders) but metadata• Items – Attributes – Relationships – Schemas –

Sub-Schemas (extensions)– Déjà vu?


Type-Based Metadata (2)

• Ex.: image schema in WinFS


Observations

• So far, all applications using these metadata are local– It is a typical semantic interoperability problem! Efficient, distributed WinFS is not for tomorrow…

• Image metadata will always be incomplete and subjective

• #images >> #peers >> #schemas

• All these formats can be formally described by a subset of Description Logics– Use them all in and in the darkness bind them!


Outline of my Project

• Objective: large-scale image retrieval framework taking advantage of metadata

• Outline– Import images– Import metadata– Extract low-level features (thanks to Lei :-)– Store everything in a common, scalable representation– Export data in a shared repository

• SQL server• P2P network (SP2 ?)

– Infer Metadata / Schema mappings locally– Cross validate mappings– Cluster peers / images vis-à-vis their subjective views


Specificities

• Different metadata models• Incompleteness of metadata (e.g., WinFS dangling

links)• Metadata sparseness• Few (but widely-used) core-classes• Many custom extensions• Many resources• Low-level representation of the resources• Embedded user feedback

Unique application


Finding Mapping Candidates (sketch)

• [T,I,U] U-Inference based on mutual information

(scalable!)

schema, metadata

Low-levelfeatures

Low-levelfeatures

features

metadata

schema, metadata

schema, metadata

metadataschema

feedback

metadata,schemas


Cross-Validating Mappings (sketch)

• [S,M] Cross-validation based on graph partitioning, semantic gossiping or SAT techniques

Ref.: Instance-based Schema Matching for Web Databases by Domain-specific Query Probing

Jiying Wang, Ji-Rong Wen, Frederick H. Lochovsky, Wei-Ying MaVLDB 2004


Conclusions

• Leveraging local metadata produced by end-users– Complex problem! Good heuristics could take years to

be developed…

• Local communications / computations– Scalability

• Hopefully, better results than keywords / low-level analyses even for simple heuristics– Take advantage of context

• Images given local semantics• Analyze the dynamics of the overall system• Objectivity vs. subjectivity of interpretation

becomes a measurable quality


References

• RDF/S: http://www.w3.org/

• XMP: http://www.adobe.com/products/xmp/

• WinFS: http://msdn.microsoft.com/Longhorn

• Chatty Web: http://lsirpeople.epfl.ch/cudre/

• Ontology Alignment: knowledgeweb D.2.2.3.– (send me an email to f-pcudre)

Documents

©2004, Philippe Cudré-Mauroux Semantic Interoperability for Global Information Systems Microsoft Research Asia 08.20.04 Philippe Cudré-Mauroux Distributed