Composing Mappings between Schemas using a Reference Ontology

1
Composing Mappings between Schemas using a
Reference Ontology
Eduard Dragut, Ramon Lawrence Iowa Database and
Emerging Applications (IDEA) Laboratory University
of Iowa eduard-dragut, ramon-lawrence_at_uiowa.edu

2
Outline

• Motivation
• Integration Approach
• Background
• Architecture Overview
• Ontological Matching
• Composing Mappings
• Global View Construction
• Experimental Results
• Future Work and Conclusions

3
Motivation

• Many organizations have pre-existing ontologies
  that are not suitable as global views but are
  suitable as reference ontologies to aid
  integration.
• Example National Cancer Institute (NCI) and
  National Insitutes of Health (NIH) have caBIG
  grid prototype which standardizes terminology
  (EVS, caDSR) and data elements in cancer domain.
• Schema-to-ontology matching requires integrators
  understand only their schema instead of all
  schemas that they may want to integrate.

4
Integration Approach
Reference Ontology
Page 4
Composing Mappings between Schemas using a
Reference Ontology - ODBASE04 - Eduard Dragut,
Ramon Lawrence
5
BackgroundOntologies and Integration

• Ontologies as the integrated, global view
• Carnot project (Collet91) with Cyc ontology
  (Lenat90)
• ONTOBROKER (Decker98), OBSERVER (Mena00)
• Tools for semi-automatically merging ontologies
• PROMPT (Noy00), Ontobuilder (Gal04)
• Use ontologies as matching/integration aids
• MOMIS (Beneventano03) using WordNet
• Indirect (Xu03), CUPID (Madhavan01), COMA (Do02)
• Matching ontologies (Doan02)
• Discovering ontologies (Madhavan03)
• Corpus-based matching

6
BackgroundModel Management

• Model management as proposed by (Bernstein03) is
  intended to allow high-level schema operations.
• Operators include Invert, Compose, Match, Merge.
• Warning Semantics of all operators are not yet
  fully defined and some of them are not completely
  automatic.
• Definitions
• A match is a semantic correspondence between
  schema elements.
• A mapping between schema elements is an
  expression that relates the elements.
• Note that most schema matching systems such as
  COMA produce matches not mappings.

7
Architecture Overview

• We assume the existence of a pre-existing
  reference ontology that has been accepted in a
  domain.
• The ontology is NOT a global view and may not
  cover the information in all schemas. It cannot
  be edited.
• Global view construction is a 3-step process
• 1) Independently match each schema to the
  ontology.
• 2) Compose schema-to-ontology matches to produce
  schema-to-schema mappings.
• 3) Merge the schema mappings to produce the
  global view.
• The challenge is to automate this as much as
  possible.

8
Benefits of Approach

• Even with manual integration there are several
  benefits to using a reference ontology
• 1) An integrator must only understand their
  schema and the ontology and not other schemas to
  be integrated.
• 2) Most validation is performed once during
  schema-to-ontology matching and not for every
  schema integrated.
• 3) Schema-to-ontology matchings can be re-used
  every time a new schema is integrated into the
  federation.
• Automation can
• 1) Help construct schema-to-ontology matchings.
• 2) Perform composition of mappings.
• 3) Build a global view from the composed mappings.

9
Automation Challenges

• There are several challenges in automating this
  process
• 1) Schema matching systems such as COMA are
  designed for simpler relational schemas.
  Ontologies must be mapped into a suitable format
  for use with COMA.
• 2) Schema-to-ontology matching is less accurate
  due to more complicated ontological structure and
  because the ontology may not model the entire
  domain or may model it differently.
• 3) Composing matchings often results in many
  false matches which must be handled.
• 4) A method for merging schemas using model
  management primitive operators is required.
• Even with these operators, Merge is not fully
  automatic.

10
BackgroundCOMA

• COMA (Do02) is a schema matching system that can
  flexibly combine different match algorithms and
  re-use match results.
• Match algorithms use names, paths, and schema
  properties in various ways.
• The mapping format between two schemas R and S is
  a triple (r,s,v) where r in R, s in S, and v is
  the similarity value in 0..1 between elements r
  and s.
• A schema in COMA is represented as a rooted
  directed acyclic graph. Schema elements are
  nodes which may be connected by links of
  different types.

11
Ontological Matching

• The first step is to convert ontologies in
  OWL/DAML format into COMAs graph representation
  format.
• Wrote a program that used the JENA parser.
• During the conversion
• 1) Explicitly converted a named relationship in
  the ontology into a node and several edges in
  graph.
• 2) Explicitly encoded attributes inherited over
  IS-A links since COMA does not support IS-A.
• After conversion, COMA would automatically
  produce a schema-to-ontology match as it would
  appear to be matching two relational schemas.

12
Converting Ontology to a Graph
13
Ontological MatchingMax versus noMax

• One challenge is what should this match look
  like?
• Two choices
• 1) Max - For each schema element, keep the best
  match with the ontology (if any).
• 2) NoMax - For each schema element, keep all the
  matches that are above the cutoff threshold.
• Since Max only generates one match, it is
  probably the best in semi-automated settings.
  NoMax will generate many matches which must be
  filtered out by the user or during composition.

14
Composing Mappings

• Schema-to-ontology mappings must be composed to
  produce direct schema-to-schema mappings.
• Since mappings carry no semantics, two objects
  are assumed to be identical if they map to the
  same ontological concept. Composition is
  performed transitively and is implemented using a
  natural join.
• That is, if element r is similar to o and o is
  similar to s, then we assume that r is similar to
  s.
• For example
• ltpostalCode,Zip,0.8gt and ltZip, postCode,0.7gt can
  be composed to yield ltpostalCode,postCode,0.75gt.
• The similarity values may be combined using
  various functions, although average is the most
  common.

15
Composition Example
16
Global View Construction

• One of the possible applications of constructing
  schema-to-schema mappings in this way is using
  them to build a global view.
• We have given a script in the paper that uses
  model management operators to compose any number
  of schema-to-ontology mappings into a single
  global view for all sources.
• Note that this algorithm is not perfect nor fully
  automatic as the mappings are not perfect and the
  Merge operator may require human intervention.

17
Global View Construction Example
18
Experimental Setup

• Matched the 5 sample order schemas CIDR, Excel,
  Noris, Paragon, and Apertum used to evaluate
  COMA. Numbered these schemas 1, 2, 3, 4, and 5.
• Created a reference ontology that models some of
  the domain (but not all of it) and is quite
  different than the schemas (uses IS-A for
  example).
• Used the matchings specified with COMA as
  ground-truth.
• Evaluation metrics
• Precision - of correct matches/ of suggested
  matches
• Recall - of correct matches returned/ total
  matches
• Overall Recall (2 - 1 / Precision)

19
Reference Order Ontology
20
Experiment 1Schema-to-Ontology Matching

• Goal Evaluate the accuracy of schema-to-ontology
  matching.
• Method
• Automatically convert ontology into COMA format
  and match each schema with ontology.
• Evaluation
• Measured the percent overlap of the schema and
  ontology. For many schemas, only 60 of their
  concepts were in the ontology.
• Evaluated the precision, recall, and overall
  measures relative to the number of matches that
  could be found.
• E.g. If overlap was 60 and recall was 50, then
  only 30 of all schema elements were matched BUT
  of all the possible matches, 50 were found.

21
Experiment 1 Results
noMax is poor for schema 5 as Buyer incorrectly
matched to ontology.
22
Experiment 2Schema-to-Schema Mappings

• Goal Determine the accuracy of producing
  schema-to-schema mappings by composing
  schema-to-ontology matchings.
• Method
• Used automatically generated schema-to-ontology
  matchings and composed them. Evaluated
  composition result against COMA answers for
  direct matching.
• Evaluated noMax and Max techniques and manual
  mappings.

23
Experiment 2 Results (Overall)
1 lt-gt 2 is poor because of Street mapping. 4
lt-gt 5 is poor because of Buyer mapping.
24
Experiment 3Improving Direct Matches

• Goal Determine if the accuracy of producing
  direct schema-to-schema mappings can be improved
  by re-using schema-to-ontology matches.
• Method
• Generate schema-to-schema mappings by composing
  schema-to-ontology matchings and then use this as
  past matching information for COMA.
• Allow COMA to perform direct match given this
  information.
• Evaluated noMax and Max techniques and manual
  mappings.

25
Experiment 3 Results (Overall)
1 lt-gt 2 is poor because of Street mapping.
26
Discussion and Conclusions

• Major findings
• 1) Schema-to-ontology mappings can be constructed
  with good accuracy (70-80 precision, 60
  recall).
• 2) The composition of schema-to-ontology
  matchings produces similar results to direct
  matching with COMA.
• 3) Max has higher precision than noMax but with
  lower recall. Max is probably best when the user
  must filter incorrect matches and always saves
  work.
• 4) It is valuable to re-use schema-to-ontology
  matchings (either automatic or manually
  constructed) to improve the accuracy of direct
  matchings.
• Major conclusion There is a benefit to building
  semi-automatic schema-to-ontology matchings for
  use in integration and global view construction.

27
Future Work and Challenges

• The major challenge is that the mappings carry no
  semantics which often results in incorrect
  matches suggested after composition.
• We are currently working on extending the
  mappings to capture semantics to avoid many of
  these cases.
• The approach is not fully automatic (nor will it
  ever be). However, most manual work is in the
  schema-to-ontology matching stage. We need
  better algorithms and tools to support this
  matching.
• Want to perform experimental evaluation on larger
  ontologies such as those from NCI.
• Issue Many ontologies are not in suitable form
  for intermediate mapping with schemas. (just
  taxonomies)

28
Composing Mappings between Schemas using a
Reference Ontology
Eduard Dragut, Ramon Lawrence Iowa Database and
Emerging Applicatons (IDEA) Laboratory University
of Iowa eduard-dragut, ramon-lawrence_at_uiowa.edu
29
Extra Slides
Extra Slides...
30
Ontology Conversion Algorithm

• 1) Each ontology concept (class) becomes a node
  in the graph.
• 2) For each property (attribute) of a class, add
  a node to the graph and connect it to its class.
• 3) Non-basetype properties (those with domain and
  range in ontology) are converted by
• 3a) Creating a node in the graph for the
  relationship.
• 3b) Adding an edge from the class domain to this
  node.
• 3c) Adding an edge from the new node to the range
  class.
• Note Do not currently support properties that
  have a domain or range that is union/intersection
  of concepts.
• 4) IS-A expanded by graph traversal.

31
Mapping Composition Challenges
Composing N1 match with 1N match results in a
cross-product
Cannot handle these cases as mappings have no
semantics.
32
Global View Construction Script
Computes Global View of N Source Schemas (with
ontology mappings)
Operator GlobalView(ArraySchemas, ArrayMappings,
O, n) // ArraySchemas stores the n schemas //
ArrayMappings stores the n schema-to-ontology
mappings 1. If n lt 0 Then Return empty
schema 2. If n 1 Then Return
ArraySchemas0 3. S1 ArraySchemas0 4. S2
ArraySchemas1 5. map1 ArrayMappings0 6.
map2 ArrayMappings1 7. lt S, map gt
GlobalView2(S1, S2, map1, map2, O) 8. For (i2
i lt n-1 i) 9. S1 S 10. map1
map 11. S2 ArraySchemasi 12. map2
ArrayMappingsi 13. lt S, map gt
GlobalView2(S1, S2, map1, map2, O) 14. end
for 15. Return lt S, map gt
33
Global View Construction Script (2)
Computes Global View of Two Source Schemas (with
ontology mappings)
Operator GlobalView2(S1, S2, O, S1_O, S2_O) 1.
S1_S2 S1_O Invert(S2_O) 2. lt M, S1_M, S2_M gt
Merge(S1, S2, S1_S2) 3. M_O Invert(S1_M)
S1_O Invert(S2_M) S2_O 4. Return lt M, M_O gt
34
Sample Order SchemaExcel XML Schema
lt?xml version"1.0"?gt ltSchema name"PurchaseOrder.
biz" xmlns"urnschemas-microsoft-comxml-data"
xmlnsdt"urnschemas-microsoft-comdatatypes"gt
ltElementType name"PurchaseOrder"
content"eltOnly"gt ltelement
type"Header"/gt ltelement type"Items"/gt lteleme
nt type"Footer"/gt ltelement type"InvoiceTo"/gt
ltelement type"DeliverTo"/gt lt/ElementTypegtltEleme
ntType name"Items" content"eltOnly"gt ltAttribut
eType name"itemCount" dttype"int"gtlt/AttributeTy
pegt ltattribute type"itemCount"/gt ltelement
type"Item" maxOccurs"" minOccurs"1"/gt lt/Eleme
ntTypegt ltElementType name"Item"
content"empty"gt ltAttributeType
name"yourPartNumber" dttype"string"gtlt/Attribute
Typegt ltAttributeType name"unitPrice"
dttype"number"gtlt/AttributeTypegt ltAttributeType
name"unitOfMeasure" dttype"string"gtlt/Attribute
Typegt ltAttributeType name"salesValue"
dttype"number"gtlt/AttributeTypegt ltAttributeType
name"quantity" dttype"number"gtlt/AttributeTypegt
ltAttributeType name"partNumber"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"partDescription" dttype"string"gtlt/Attribu
teTypegt ltAttributeType name"itemNumber"
dttype"int"gtlt/AttributeTypegt
35
Sample Order SchemaExcel XML Schema (2)
ltattribute type"itemNumber"/gt ltattribute
type"yourPartNumber"/gt ltattribute
type"partNumber"/gt ltattribute
type"partDescription"/gt ltattribute
type"quantity"/gt ltattribute type"unitOfMeasure
"/gt ltattribute type"unitPrice"/gt ltattribute
type"salesValue"/gt lt/ElementTypegt ltElementType
name"InvoiceTo" content"eltOnly"gt ltelement
type"Contact"/gt ltelement type"Address"/gt lt/E
lementTypegt ltElementType name"Header"
content"eltOnly"gt ltAttributeType
name"yourAccountCode" dttype"string"gtlt/Attribut
eTypegt ltAttributeType name"ourAccountCode"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"orderNum" dttype"string"gtlt/AttributeTypegt
ltAttributeType name"orderDate"
dttype"date"gtlt/AttributeTypegt ltattribute
type"orderNum"/gt ltattribute type"orderDate"/gt
ltattribute type"ourAccountCode"/gt ltattribute
type"yourAccountCode"/gt ltelement
type"Contact"/gt lt/ElementTypegt
36
Sample Order SchemaExcel XML Schema (3)
ltElementType name"Footer" content"empty"gt
ltAttributeType name"totalValue"
dttype"number"gtlt/AttributeTypegt ltattribute
type"totalValue"/gt lt/ElementTypegt ltElementType
name"DeliverTo" content"eltOnly"gt ltelement
type"Contact"/gt ltelement type"Address"/gt lt/E
lementTypegt ltElementType name"Contact"
content"empty"gt ltAttributeType
name"telephone" dttype"string"gtlt/AttributeTypegt
ltAttributeType name"e-mail"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"contactName" dttype"string"gtlt/AttributeTy
pegt ltAttributeType name"companyName"
dttype"string"gtlt/AttributeTypegt ltattribute
type"contactName"/gt ltattribute
type"companyName"/gt ltattribute
type"e-mail"/gt ltattribute type"telephone"/gt lt
/ElementTypegt
37
Sample Order SchemaExcel XML Schema (4)
ltElementType name"Address" content"empty"gt
ltAttributeType name"street4"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"street3" dttype"string"gtlt/AttributeTypegt
ltAttributeType name"street2"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"street1" dttype"string"gtlt/AttributeTypegt
ltAttributeType name"stateProvince"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"postalCode" dttype"string"gtlt/AttributeTyp
egt ltAttributeType name"country"
dttype"string"gtlt/AttributeTypegt ltAttributeType
name"city" dttype"string"gtlt/AttributeTypegt lt
attribute type"street1"/gt ltattribute
type"street2"/gt ltattribute type"street3"/gt lt
attribute type"street4"/gt ltattribute
type"city"/gt ltattribute type"stateProvince"/gt
ltattribute type"postalCode"/gt ltattribute
type"country"/gt lt/ElementTypegt lt/Schemagt
38
Experiment 2 Precision
39
Experiment 2 Recall
40
Experiment 3 Results (Precision)
41
Experiment 3 Results (Recall)