Learning to Map between
Structured Representations of Data
AnHai Doan
Database & Data Mining Group
University of Washington, Seattle
Spring 2002
Data Integration Challenge
Find houses with
2 bedrooms
priced under
200K
New faculty
member
realestate.com
homeseekers.com
homes.com
2
Architecture of Data Integration System
Find houses with 2 bedrooms
priced under 200K
mediated schema
source schema 1
realestate.com
source schema 2
homeseekers.com
source schema 3
homes.com
3
Semantic Mappings between Schemas
Mediated-schema
price agent-name
1-1 mapping
homes.com
listed-price
320K
240K
contact-name
Jane Brown
Mike Smith
address
complex mapping
city
state
Seattle WA
Miami FL
4
Schema Matching is Ubiquitous!


Fundamental problem in numerous applications
Databases
–
–
–
–
–
–
–

data integration
data translation
schema/view integration
data warehousing
semantic query processing
model management
peer data management
AI
– knowledge bases, ontology merging, information gathering agents, ...

Web
– e-commerce
– marking up data using ontologies (Semantic Web)
5
Why Schema Matching is Difficult

Schema & data never fully capture semantics!
– not adequately documented

Must rely on clues in schema & data
– using names, structures, types, data values, etc.

Such clues can be unreliable
– same names => different entities: area => location or square-feet
– different names => same entity:
area & address => location

Intended semantics can be subjective
– house-style = house-description?

Cannot be fully automated, needs user feedback!
6
Current State of Affairs

Finding semantic mappings is now a key bottleneck!
– largely done by hand
– labor intensive & error prone
– data integration at GTE [Li&Clifton, 2000]
– 40 databases, 27000 elements, estimated time: 12 years

Will only be exacerbated
– data sharing becomes pervasive
– translation of legacy data


Need semi-automatic approaches to scale up!
Many current research projects
– Databases: IBM Almaden, Microsoft Research, BYU, George Mason,
U of Leipzig, ...
– AI: Stanford, Karlsruhe University, NEC Japan, ...
7
Goals and Contributions

Vision for schema-matching tools
–
–
–
–

learn from previous matching activities
exploit multiple types of information
incorporate domain integrity constraints
handle user feedback
My contributions:
solution for semi-automatic schema matching
–
–
–
–
can match relational schemas, DTDs, ontologies, ...
discovers both 1-1 & complex mappings
highly modular & extensible
achieves high matching accuracy (66 -- 97%) on real-world data
8
Road Map


Introduction
Schema matching
[SIGMOD-01]
– 1-1 mappings for data integration
– LSD (Learning Source Description) system
– learns from previous matching activities
– employs multi-strategy learning
– exploits domain constraints & user feedback



Creating complex mappings
Ontology matching [WWW-02]
Conclusions
[Tech. Report-02]
9
Schema Matching for Data Integration:
the LSD Approach
Suppose user wants to integrate 100 data sources
1. User
– manually creates mappings for a few sources, say 3
– shows LSD these mappings
2. LSD learns from the mappings
3. LSD predicts mappings for remaining 97 sources
10
Learning from the Manual Mappings
price
Mediated schema
agent-name agent-phone office-phone
description
listed-price contact-name contact-phone office comments
If “office”
occurs in name
=> office-phone
Schema of realestate.com
realestate.com
listed-price
contact-name contact-phone
$250K
$320K
James Smith
Mike Doan
$350K
$230K
contact-agent
comments
(305) 729 0831 (305) 616 1822 Fantastic house
(617) 253 1429 (617) 112 2315 Great location
homes.com
sold-at
office
extra-info
(206) 634 9435 Beautiful yard
(617) 335 4243 Close to Seattle
If “fantastic” & “great”
occur frequently in
data instances
=> description
11
Must Exploit Multiple Types of Information!
Mediated schema
price
agent-name agent-phone office-phone
description
listed-price contact-name contact-phone office comments
If “office”
occurs in name
=> office-phone
Schema of realestate.com
realestate.com
listed-price
contact-name contact-phone
$250K
$320K
James Smith
Mike Doan
$350K
$230K
contact-agent
comments
(305) 729 0831 (305) 616 1822 Fantastic house
(617) 253 1429 (617) 112 2315 Great location
homes.com
sold-at
office
extra-info
(206) 634 9435 Beautiful yard
(617) 335 4243 Close to Seattle
If “fantastic” & “great”
occur frequently in
data instances
=> description
12
Multi-Strategy Learning

Use a set of base learners
– each exploits well certain types of information

To match a schema element of a new source
– apply base learners
– combine their predictions using a meta-learner

Meta-learner
– uses training sources to measure base learner accuracy
– weighs each learner based on its accuracy
13
Base Learners

Training
Object
Training
examples



Matching
Name Learner
(X1,C1)
(X2,C2)
...
(Xm,Cm)
Observed label
X
labels weighted by confidence score
Classification model
(hypothesis)
– training:
(“location”, address)
(“contact name”, name)
– matching:
agent-name => (name,0.7),(phone,0.3)
Naive Bayes Learner
– training:
(“Seattle, WA”,address)
(“250K”,price)
– matching:
“Kent, WA”
=> (address,0.8),(name,0.2)
14
The LSD Architecture
Training Phase
Matching Phase
Mediated schema
Source schemas
Base-Learner1
Hypothesis1
Training data
for base learners
Base-Learnerk
Hypothesisk
Base-Learner1 .... Base-Learnerk
Meta-Learner
Predictions for instances
Prediction Combiner
Domain
constraints Predictions for elements
Constraint Handler
Meta-Learner
Weights for
Base Learners
Mappings
15
Training the Base Learners
Mediated schema
address
price agent-name agent-phone office-phone
description
realestate.com
location
price contact-name contact-phone
office
comments
Miami, FL $250K James Smith (305) 729 0831 (305) 616 1822 Fantastic house
Boston, MA $320K Mike Doan (617) 253 1429 (617) 112 2315 Great location
Name Learner
Naive Bayes Learner
(“location”, address)
(“price”, price)
(“contact name”, agent-name)
(“contact phone”, agent-phone)
(“office”, office-phone)
(“comments”, description)
(“Miami, FL”, address)
(“$250K”, price)
(“James Smith”, agent-name)
(“(305) 729 0831”, agent-phone)
(“(305) 616 1822”, office-phone)
(“Fantastic house”, description)
(“Boston,MA”, address)
16
Meta-Learner: Stacking
[Wolpert 92,Ting&Witten99]

Training
–
–
–
–

uses training data to learn weights
one for each (base-learner,mediated-schema element) pair
weight (Name-Learner,address) = 0.2
weight (Naive-Bayes,address) = 0.8
Matching: combine predictions of base learners
– computes weighted average of base-learner confidence scores
area
Seattle, WA
Kent, WA
Bend, OR
Name Learner
Naive Bayes
(address,0.4)
(address,0.9)
Meta-Learner
(address, 0.4*0.2 + 0.9*0.8 = 0.8)
17
The LSD Architecture
Training Phase
Matching Phase
Mediated schema
Source schemas
Base-Learner1
Hypothesis1
Training data
for base learners
Base-Learnerk
Hypothesisk
Base-Learner1 .... Base-Learnerk
Meta-Learner
Predictions for instances
Prediction Combiner
Domain
constraints Predictions for elements
Constraint Handler
Meta-Learner
Weights for
Base Learners
Mappings
18
Applying the Learners
homes.com schema
area
sold-at contact-agent
area
Seattle, WA
Kent, WA
Bend, OR
extra-info
Name Learner
Naive Bayes
Name Learner
Naive Bayes
Meta-Learner
Meta-Learner
(address,0.8), (description,0.2)
(address,0.6), (description,0.4)
(address,0.7), (description,0.3)
Prediction-Combiner
(address,0.7), (description,0.3)
homes.com
sold-at
contact-agent
extra-info
(price,0.9), (agent-phone,0.1)
(agent-phone,0.9), (description,0.1)
(address,0.6), (description,0.4)
19
Domain Constraints



Encode user knowledge about domain
Specified by examining mediated schema
Examples
– at most one source-schema element can match address
– if a source-schema element matches house-id then it is a key
– avg-value(price) > avg-value(num-baths)

Given a mapping combination
– can verify if it satisfies a given constraint
area:
sold-at:
contact-agent:
extra-info:
address
price
agent-phone
address
20
The Constraint Handler
Predictions from Prediction Combiner
Domain Constraints
area:
(address,0.7), (description,0.3)
sold-at:
(price,0.9), (agent-phone,0.1)
contact-agent: (agent-phone,0.9), (description,0.1)
extra-info:
(address,0.6), (description,0.4)
At most one element
matches address
area:
sold-at:
contact-agent:
extra-info:



address
price
agent-phone
address
0.7
0.9
0.9
0.6
0.3402
area:
sold-at:
contact-agent:
extra-info:
address
price
agent-phone
description
0.3
0.1
0.1
0.4
0.0012
0.7
0.9
0.9
0.4
0.2268
Searches space of mapping combinations efficiently
Can handle arbitrary constraints
Also used to incorporate user feedback
– sold-at does not match price
21
The Current LSD System

Can also handle data in XML format
– matches XML DTDs

Base learners
– Naive Bayes [Duda&Hart-93, Domingos&Pazzani-97]
– exploits frequencies of words & symbols
– WHIRL Nearest-Neighbor Classifier [Cohen&Hirsh KDD-98]
– employs information-retrieval similarity metric
– Name Learner [SIGMOD-01]
– matches elements based on their names
– County-Name Recognizer [SIGMOD-01]
– stores all U.S. county names
– XML Learner [SIGMOD-01]
– exploits hierarchical structure of XML data
22
Empirical Evaluation

Four domains
– Real Estate I & II, Course Offerings, Faculty Listings

For each domain
–
–
–
–

created mediated schema & domain constraints
chose five sources
extracted & converted data into XML
mediated schemas: 14 - 66 elements, source schemas: 13 - 48
Ten runs for each domain, in each run:
– manually provided 1-1 mappings for 3 sources
– asked LSD to propose mappings for remaining 2 sources
– accuracy = % of 1-1 mappings correctly identified
23
Average Matching Acccuracy (%)
High Matching Accuracy
100
90
80
70
60
50
40
30
20
10
0
Real Estate I
Real Estate II
Course
Offerings
Faculty
Listings
LSD’s accuracy:
71 - 92%
Best single base learner: 42 - 72%
+ Meta-learner:
+ 5 - 22%
+ Constraint handler:
+ 7 - 13%
+ XML learner:
+ 0.8 - 6%
24
Average matching accuracy (%)
Contribution of Schema vs. Data
100
90
80
70
60
50
40
30
20
10
0
Real Estate I
Real Estate II
Course Offerings Faculty Listings
LSD with only schema info.
LSD with only data info.
Complete LSD
More experiments in [Doan et al. SIGMOD-01]
25
LSD Summary

LSD
– learns from previous matching activities
– exploits multiple types of information
– by employing multi-strategy learning
– incorporates domain constraints & user feedback
– achieves high matching accuracy

LSD focuses on 1-1 mappings
 Next challenge: discover more complex mappings!
– COMAP (Complex Mapping) system
26
The COMAP Approach
Mediated-schema
price num-baths
address
homes.com
listed-price
320K
240K

agent-id
full-baths half-baths city
53211
11578
2
1
1
1
Seattle
Miami
zipcode
98105
23591
For each mediated-schema element
– searches space of all mappings
– finds a small set of likely mapping candidates
– uses LSD to evaluate them

To search efficiently
– employs a specialized searcher for each element type
– Text Searcher, Numeric Searcher, Category Searcher, ...
27
The COMAP Architecture [Doan et al., 02]
Mediated schema
Searcher1
Source schema + data
Searcher2
Searcherk
Mapping candidates
Base-Learner1 .... Base-Learnerk
Meta-Learner
Domain
constraints
Prediction Combiner
Constraint Handler
Mappings
LSD
28
An Example: Text Searcher


Beam search in space of all concatenation mappings
Example: find mapping candidates for address
Mediated-schema
homes.com
listed-price
320K
240K
price num-baths
agent-id full-baths half-baths city
zipcode
532a
115c
98105
23591
concat(agent-id,city)
532a Seattle
115c Miami

address
2
1
1
1
Seattle
Miami
concat(agent-id,zipcode)
532a 98105
115c 23591
concat(city,zipcode)
Seattle 98105
Miami 23591
Best mapping candidates for address
– (agent-id,0.7), (concat(agent-id,city),0.75), (concat(city,zipcode),0.9)
29
Empirical Evaluation

Current COMAP system
– eight searchers

Three real-world domains
– in real estate & product inventory
– mediated schema: 6 -- 26 elements, source schema: 16 -- 31

Accuracy: 62 -- 97%
 Sample discovered mappings
– agent-name = concat(first-name,last-name)
– area
= building-area / 43560
– discount-cost = (unit-price * quantity) * (1 - discount)
30
Road Map


Introduction
Schema matching
– LSD system

Creating complex mappings
– COMAP system

Ontology matching
– GLUE system

Conclusions
31
Ontology Matching

Increasingly critical for
– knowledge bases, Semantic Web

An ontology
– concepts organized into a taxonomy tree
– each concept has
– a set of attributes
– a set of instances
Undergrad
Courses
– relations among concepts

CS Dept. US
Entity
Grad
Courses
Faculty
Matching
– concepts
– attributes
– relations
People
Assistant
Professor
Associate
Professor
Staff
Professor
name: Mike Burns
degree: Ph.D.
32
Matching Taxonomies of Concepts
CS Dept. US
CS Dept. Australia
Entity
Undergrad
Courses
Grad
Courses
Entity
People
Faculty
Assistant
Professor
Associate
Professor
Courses
Staff
Professor
Lecturer
Staff
Academic Staff
Technical Staff
Senior
Lecturer
Professor
33
Constraints in Taxonomy Matching

Domain-dependent
– at most one node matches department-chair
– a node that matches professor can not be a child of a node
that matches assistant-professor

Domain-independent
– two nodes match if parents & children match
– if all children of X matches Y, then X also matches Y
– Variations have been exploited in many restricted settings
[Melnik&Garcia-Molina,ICDE-02], [Milo&Zohar,VLDB-98],
[Noy et al., IJCAI-01], [Madhavan et al., VLDB-01]

Challenge: find a general & efficient approach
34
Solution: Relaxation Labeling

Relaxation labeling [Hummel&Zucker, 83]
–
–
–
–

applied to graph labeling in vision, NLP, hypertext classification
finds best label assignment, given a set of constraints
starts with initial label assignment
iteratively improves labels, using constraints
Standard relax. labeling not applicable
– extended it in many ways [Doan et al., W W W-02]

Experiments
–
–
–
–
three real-world domains in course catalog & company listings
30 -- 300 nodes / taxonomy
accuracy 66 -- 97% vs. 52 -- 83% of best base learner
relaxation labeling very fast (under few seconds)
35
Related Work
Hand-crafted rules
Exploit schema
1-1 mapping
TRANSCM [Milo&Zohar98]
ARTEMIS [Castano&Antonellis99]
[Palopoli et al. 98]
CUPID
[Madhavan et al. 01]
PROMPT [Noy et al. 00]
Rules
Exploit data
1-1 + complex mapping
CLIO [Miller et. al., 00]
[Yan et al. 01]
Single learner
Exploit data
1-1 mapping
SEMINT [Li&Clifton94]
ILA [Perkowitz&Etzioni95]
DELTA [Clifton et al. 97]
Learners + rules, use multi-strategy learning
Exploit schema + data
1-1 + complex mapping
Exploit domain constraints
LSD [Doan et al., SIGMOD-01]
COMAP [Doan et al. 2002, submitted]
GLUE [Doan et al., WWW-02]
36
Future Work

Learning source descriptions
– formal semantics for mapping
– query capabilities, source schema, scope, reliability of data, ...

Dealing with changes in source description
 Matching objects across sources
 More sophisticated user feedback

Focus on distributed information management systems
– data integration, web-service integration, peer data management
– goal: significantly reduce complexity of construction & maintenance
37
Conclusions


Efficiently creating semantic mappings is critical
Developed solution for semi-automatic schema matching
–
–
–
–
–

learns from previous matching activities
can match relational schemas, DTDs, ontologies, ...
discovers both 1-1 & complex mappings
highly modular & extensible
achieves high matching accuracy
Made contributions to machine learning
– developed novel method to classify XML data
– extended relaxation labeling
38
Backup Slides
39
Training the Meta-Learner

For address
Extracted XML Instances
<location> Miami, FL</>
<listed-price> $250,000</>
<area> Seattle, WA </>
<house-addr>Kent, WA</>
<num-baths>3</>
...
Name Learner Naive Bayes
0.5
0.4
0.3
0.6
0.3
...
0.8
0.3
0.9
0.8
0.3
...
True Predictions
1
0
1
1
0
...
Least-Squares
Linear Regression
Weight(Name-Learner,address) = 0.1
Weight(Naive-Bayes,address) = 0.9
40
Average matching accuracy (%)
Sensitivity to Amount of Available Data
100
90
80
70
60
50
40
0
100
200
300
400
500
Number of data listings per source (Real Estate I)
41
Average Matching Acccuracy (%)
Contribution of Each Component
100
80
60
40
20
0
Real Estate I
Course Offerings
Without Name Learner
Without Naive Bayes
Without Whirl Learner
Without Constraint Handler
The complete LSD system
Faculty Listings
Real Estate II
42
Exploiting Hierarchical Structure

Existing learners flatten out all structures
<contact>
<name> Gail Murphy </name>
<firm> MAX Realtors </firm>
</contact>

<description>
Victorian house with a view. Name your price!
To see it, contact Gail Murphy at MAX Realtors.
</description>
Developed XML learner
– similar to the Naive Bayes learner
– input instance = bag of tokens
– differs in one crucial aspect
– consider not only text tokens, but also structure tokens
43
Reasons for Incorrect Matchings

Unfamiliarity
– suburb
– solution: add a suburb-name recognizer

Insufficient information
– correctly identified general type, failed to pinpoint exact type
– agent-name
phone
Richard Smith
(206) 234 5412
– solution: add a proximity learner

Subjectivity
– house-style = description?
Victorian
Beautiful neo-gothic house
Mexican
Great location
44
Evaluate Mapping Candidates

For address, Text Searcher returns
– (agent-id,0.7)
– (concat(agent-id,city),0.8)
– (concat(city,zipcode),0.75)

Employ multi-strategy learning to evaluate mappings
 Example: (concat(agent-id,city),0.8)
– Naive Bayes Learner: 0.8
– Name Learner: “address” vs. “agent id city” 0.3
– Meta-Learner: 0.8 * 0.7 + 0.3 * 0.3 = 0.65

Meta-Learner returns
– (agent-id,0.59)
– (concat(agent-id,city),0.65)
– (concat(city,zipcode),0.70)
45
Relaxation Labeling

Applied to similar problems in
– vision, NLP, hypertext classification
People
Dept Australia
Courses
Courses
Acad. Staff
Faculty
Dept U.S.
Courses
Staff
Tech. Staff
Staff
Courses
People
Faculty
Staff
46
Relaxation Labeling for Taxonomy Matching

Must define
– neighborhood of a node
– k features of neighborhood
– how to combine influence of features
– P( N  L | f1 , f 2 ,..., f k )

Algorithm
– init:
– loop:
for each pair <N,L>, compute P( N  L | )
for each pair <N,L>, re-compute
P( N  L | )   P( N  L | M , ).P( M | )
M
Staff = People
Acad. Staff: Faculty
Tech. Staff: Staff
Neighborhood configuration
47
Relaxation Labeling for Taxonomy Matching

Huge number of neighborhood configurations!
– typically neighborhood = immediate nodes
– here neighborhood can be entire graph
100 nodes, 10 labels => 10100 configurations

Solution
– label abstraction + dynamic programming
– guarantee quadratic time for a broad range of domain constraints

Empirical evaluation
–
–
–
–
–
GLUE system [Doan et. al., WWW-02]
three real-world domains
30 -- 300 nodes / taxonomy
high accuracy 66 -- 97% vs. 52 -- 83% of best base learner
relaxation labeling very fast, finished in several seconds
48