Challenges to Developing a Syntax to Capture Metadata Mapping within XMDR

1
Challenges to Developing a Syntax to Capture
Metadata Mapping within XMDR

• (or A Social Science Data Expert Meets
  Semantics)
• Fredric Gey
• UC Berkeley
• XMDR Project Meeting
• May, 2007
• UC Berkeley Faculty Club

2
XMDR Mapping Capture (outline)

• In January 2007 proposed strawman syntax for XMDR
  to capture mapping between classification systems
• Heavily influenced by the SIC NAICS content
  loading
• Very naïve compared to real-world challenges
• Conceptual framework, example problems
• Drawn principally from ontology mapping research
• Discussion points for future development

3
Metadata/Ontology Mapping

• Ontology mapping is now a substantial research
  area
• Driven by need to integrate disparate ontologies
  produced by different organizations or different
  units of one organization
• Commercial and research tools under development
• Tools generally solve pieces of the mapping
  problem
• Five workshops on the subject OM-2006 was held
  in Georgia Nov 2006, OM-2007 will be in Busan,
  Korea this November
• Evaluating mapping tools is also being researched
• Can you compare apples and oranges (or rather
  apple processing equipment versus orange
  processing equipment)?
• OM Ontology Alignment Initiative (OM 2004, 2005,
  2006)
• Elucidates problems and generates evaluation
  frameworks
• Dimensions of mapping challenge pertinent to XMDR

4
Relationship to Schema Mapping

• Schema mapping is closely related to Ontology
  mapping
• Schema mapping researched in the database
  community
• Driven by need to integrate different databases
  produced by different organizations or different
  units of one organization
• Essential for Federated databases where database
  administration is under local (distributed)
  control
• Simple example person name
• DB1 Name type text
• DB2 Lastname, Firstname, Middleinitial, Title
  (e.g. Dr, Mr, Ms), Suffix (e.g. Jr, III)

5
Conceptual Context of Mapping

• Mapping (also known as ontology alignment) can
  consist of
• Integration of multiple component ontologies into
  a unified ontology
• Alignment between part of a global ontology to an
  external local ontology
• Alignment between different parts of global
  ontologies
• Alignment between special subject areas having
  overlapping instances (e.g. medical literature to
  computational biology literature)
• Mapping tools (Noy Musen 2002) can be
• Partial not global
• Be instance-based or not (for common or
  non-overlapping instances)
• Be lexical-based (i.e. utilize lexical clues to
  infer equivalences)
• May only deal with class hierarchies (not
  instances)
• May only produce articulation rules between
  ontologies

6
Strawman Syntax for XMDR mapping from an Old
Classification to a New Classification
ltlgRelassociation association"mapsTo"
forwardName"mapsTo" ltreverseName"mappe
dFromgt lttargetCodingScheme"coding_scheme_name"gt
-ltlgRelsourceConcept sourceConcept"source_concep
t_code"gt -ltlgReltargetConcept targetConcept"
target_concept_code "gt   ltlgRelassociationQualif
ication associationQualifier"exact
almost_exact approx" /gt
ltlgRelassociationQualification
MappingType"semantic statistical" /gt
ltlgRelassociationQualification
MapsToDegree"fraction" /gt
ltlgRelassociationQualification
MapsFromDegree"fraction" /gt
ltlgRelassociationQualification
MapsToThreshold"percent" /gt lt/lgReltargetConcept
gt  lt/lgRelsourceConceptgtlt/lgRelassociationgt
7
Strawman Syntax Definitions of Qualifiers

• MappingType
• Semantic meaning of the source concept is
  aligned with the meaning of the target concept
• Statistical mapping is done using a common
  database indexed/classified by both source and
  target concept codes
• ltassociationQualifier"exact almost_exact
  approx" /gt
• exact - mapping is 1-1
• Almost_exact - for statistical mapping, mapping
  is almost exact within some e (epsilon) percent
  difference as a MapsToThreshold
• Approx mapping is inexact with overlaps between
  concepts
• Degree of (statistical) mapping
• ltlgRelassociationQualification
  MapsToDegree"fraction1" /gt
• ltlgRelassociationQualification
  MapsFromDegree"fraction2" /gt
• i.e. a fraction2 of the source concept is
  represented by fraction1 of the target concept

8
Strawman Syntax Additional Qualifiers?

• MappingMode
• Automatic meaning of the source concept is
  automatically mapped between source and target
  concepts by some software
• Manual mapping is done with human editors
• Observation Semantic mapping could be automatic
  if done by NLP software which compares the
  textual definitions of source and concept codes
• ReferenceDatabase - the detailed database which
  has been utilized for the statistical mapping
  (e.g. 1997 Economic Census company establishment
  for SIC and NAICS)
• Your qualifier goes here

9
SIC to NAICS Matching(example for Mineral
Industries)
47
10
NAICS to SIC Matching

• Important for historical data comparison and
  development of time series

11
NAICS to SIC Matching

• Mappings are
• One to one (rarely)
• Many to one (sometimes)
• Many to many (usually)
• Census Bureau supplies
• Comparable statistics for 1997 Economics Census
• Downloadable code files with bridge
• Open Issue
• Can statistical allocation between code sets be
  captured?
• I.e. NAICS1 ? SIC1 (.35) ?SIC2(.65)

12
What are the Limitations/Assumptions of the
Strawman Proposal?

• It assumes the mapping is instance based
• Good enough for classification schemes like SIC
  and NAICS where the instances (reporting of
  individual business establishment statistics) are
  identical
• It only provides for individual
  concept-to-concept equivalence
• It assumes the structure of the source and target
  will provide context of relationships between
  (exact or amorphous) groups of concepts
• It does not capture the possibility of large
  numbers of non-overlapping concepts when the
  conceptual overlap between two ontologies is
  small
• It does not capture structural inconsistency
• It does not capture temporal drift and evolution
• It does not capture that corresponding concepts
  may be considerably more detailed (have more
  attributes) in a particular ontology

13
Ontology Integration

• Multiple component ontologies are integrated into
  a single global ontology
• Oi n Ok ? (null set), or
• Oi n Ok e
• Does XMDR need to concern itself with this case?

Oglobal
14
Ontology Structural Inconsistency (thanks Mala)

• How do we capture that two different ontologies
  may have the identical concept at a different
  structural positions in their hierarchies/directed
  graphs?

O2
O1
C1m
C11
C1m
  Computers                          
                                   
Computing Machines         /      \ 
                                            
                         /   
\         /         \              
                                   
                              /       \      
software  hardware                        
                     Internet     
Standalone     /  \           
  / \                                     
        / \                   /
\Internet Standalone Internet Standalone         
      Software Hardware Software Hardware
15
Ontology Concept Subset Mapping

• How do we capture that subsets of concepts
  between two different ontologies are equivalent
  rather than individual concepts?

O1
O2
C1k C2r
C1m C2s
16
Ontology Overlap/Non-overlap

• How do we capture that two different ontologies
  each have large numbers of concepts with no
  corresponding concept in the other ontology?
• 0(O1 n O2) ltlt 0(O1 n O2)
• Is this important?

O2
C1k ? C2r
C1m C2s
17
Ontology Granularity

• How do we capture that concepts in one ontology
  may have significantly greater detail than in
  another ontology
• In terms of subconcepts
• In terms of concept attributes/properties
• Omega is light on geographic places compared to
  GNS

GNS (Geographic Names Server)
Omega (geography)
18
Ontology Granularity (Omega)

• Omega is light (and inconsistent) on geographic
  places compared to GNIS

19
Ontology Granularity (GNS)

• GeoNames Server search of populated places in
  Finland
• Tampere, Finland is described by
• Feature type PPL (populated place)
• Latitude/Longitude
• Name variants
• http//en.wikipedia.org/wiki/Tampere

20
Discussion Issues on XMDR Mappings Syntaxes

• Beyond individual concept mapping
• Individual concept to set of concepts
• Between sets or clusters of equivalent concepts
• Capturing extent of non-overlap
• Granularity differences between equivalent
  source and target concepts
• Capturing structure in the mapping syntax
• Dealing with structural mismatches
• Can we borrow from the mathematical literature on
  topological mapping between directed graph
  structures?