1. Learning to Map Between Schemas Ontologies
Alon Halevy
University of Washington
Joint work with Anhai Doan and Pedro Domingos

2. Agenda Ontology mapping is a key problem in many applications: LSD:
Data integration
Semantic web
Knowledge management
E-commerce
LSD:
Solution that uses multi-strategy learning.
We’ve started with schema matching (I.e., very simple ontologies)
Currently extending to more expressive ontologies.
Experiments show the approach is very promising!

3. The Structure Mapping Problem
Types of structures:
Database schemas, XML DTDs, ontologies, …,
Input:
Two (or more) structures, S1 and S2
Data instances for S1 and S2
Background knowledge
Output:
A mapping between S1 and S2
Should enable translating between data instances.
Semantics of mapping?

4. Semantic Mappings between Schemas
Source schemas = XML DTDs
house
address
contact-info
num-baths
agent-name agent-phone
1-1 mapping
non 1-1 mapping
house
Points:
0) Sources export data in XML
1) Mediated & source schemas are represented with XML DTDs
2) Describe different types of mappings
1-1 mappings
more complex mappings
3) We focus on 1-1 mappings
Note: when defining the schema-matching problem, say clearly that we are
*given* the two schemas. All we need to do is to find the mappings.
Do not talk about handicap-equipped => amenities. Not enough time. Not important.
location contact
full-baths
half-baths
name phone

5. Motivation Database schema integration
A problem as old as databases themselves.
database merging, data warehouses, data migration
Data integration / information gathering agents
On the WWW, in enterprises, large science projects
Model management:
Model matching: key operator in an algebra where models and mappings are first-class objects.
See [Bernstein et al., 2000] for more.
The Semantic Web
Ontology mapping.
System interoperability
E-services, application integration, B2B applications, …,

6. Desiderata from Proposed Solutions
Accuracy, efficiency, ease of use.
Realistic expectations:
Unlikely to be fully automated. Need user in the loop.
Some notion of semantics for mappings.
Extensibility:
Solution should exploit additional background knowledge.
“Memory”, knowledge reuse:
System should exploit previous manual or automatically generated matchings.
Key idea behind LSD.

7. LSD Overview L(earning) S(ource) D(escriptions)
Problem: generating semantic mappings between mediated schema and a large set of data source schemas.
Key idea: generate the first mappings manually, and learn from them to generate the rest.
Technique: multi-strategy learning (extensible!)
Step 1:
[SIGMOD, 2001]: 1-1 mappings between XML DTDs.
Current focus:
Complex mappings
Ontology mapping.

8. Outline Overview of structure mapping
Data integration and source mappings
LSD architecture and details
Experimental results
Current work.

9. Data Integration Find houses with four bathrooms priced under $500,000
mediated schema
Query reformulation
and optimization.
source schema 1
source schema 2
source schema 3
wrappers
realestate.com
homeseekers.com
homes.com
Applications: WWW, enterprises, science projects
Techniques: virtual data integration, warehousing, custom code.

10. Semantic Mappings between Schemas
Source schemas = XML DTDs
house
address
contact-info
num-baths
agent-name agent-phone
1-1 mapping
non 1-1 mapping
house
Points:
0) Sources export data in XML
1) Mediated & source schemas are represented with XML DTDs
2) Describe different types of mappings
1-1 mappings
more complex mappings
3) We focus on 1-1 mappings
Note: when defining the schema-matching problem, say clearly that we are
*given* the two schemas. All we need to do is to find the mappings.
Do not talk about handicap-equipped => amenities. Not enough time. Not important.
location contact
full-baths
half-baths
name phone

11. Semantics (preliminary)
Semantics of mappings has received no attention.
Semantics of 1-1 mappings –
Given:
R(A1,…,An) and S(B1,…,Bm)
1-1 mappings (Ai,Bj)
Then, we postulate the existence of a relation W, s.t.:
P (C1,…,Ck) (W) = P (A1,…,Ak) (R) ,
P (C1,…,Ck) (W) = P (B1,…,Bk) (S) ,
W also includes the unmatched attributes of R and S.
In English: R and S are projections on some universal relation W, and the mappings specify the projection variables and correspondences.

12. Why Matching is Difficult
Aims to identify same real-world entity
using names, structures, types, data values, etc
Schemas represent same entity differently
different names => same entity:
area & address => location
same names => different entities:
area => location or square-feet
Schema & data never fully capture semantics!
not adequately documented, not sufficiently expressive
Intended semantics is typically subjective!
IBM Almaden Lab = IBM?
Cannot be fully automated. Often hard for humans. Committees are required!

13. Current State of Affairs
Finding semantic mappings is now the bottleneck!
largely done by hand
labor intensive & error prone
GTE: 4 hours/element for 27,000 elements [Li&Clifton00]
Will only be exacerbated
data sharing & XML become pervasive
proliferation of DTDs
translation of legacy data
reconciling ontologies on semantic web
Need semi-automatic approaches to scale up!
Points: -- what’s the state of research in data integration today?
-- this and this are well-understood
-- schema matching is now the bottleneck => need to automate

14. Outline Overview of structure mapping
Data integration and source mappings
LSD architecture and details
Experimental results
Current work.

15. The LSD Approach
User manually maps a few data sources to the mediated schema.
LSD learns from the mappings, and proposes mappings for the rest of the sources.
Several types of knowledge are used in learning:
Schema elements, e.g., attribute names
Data elements: ranges, formats, word frequencies, value frequencies, length of texts.
Proximity of attributes
Functional dependencies, number of attribute occurrences.
One learner does not fit all. Use multiple learners and combine with meta-learner.

16. Example Mediated schema Learned hypotheses Schema of realestate.com
address price agent-phone description
location listed-price phone comments
Learned hypotheses
Schema of realestate.com
If “phone” occurs in the name => agent-phone
location
Miami, FL
Boston, MA
...
listed-price
$250,000
$110,000
...
phone
(305)
(617)
...
comments
Fantastic house
Great location
...
realestate.com
If “fantastic” & “great” occur frequently in data values =>
description
Points:
1) Introduce our approach.
2) We do not manually map the schemas of all sources to mediated schema.
The goal is to manually mark up only a few sources, and be able to learn from the marked up sources to successfully propose mappings for subsequent sources.
3) Once the markup is done, there are many different types of information to learn from.
homes.com
price
$550,000
$320,000
...
contact-phone
(278)
(617)
...
extra-info
Beautiful yard
Great beach
...

17. Multi-Strategy Learning
Use a set of base learners:
Name learner, Naïve Bayes, Whirl, XML learner
And a set of recognizers:
County name, zip code, phone numbers.
Each base learner produces a prediction weighted by confidence score.
Combine base learners with a meta-learner, using stacking.

18. Base Learners Name Learner Naive Bayes Learner [Domingos&Pazzani 97]
(contact-info,office-address)
(contact-info,office-address)
(contact,agent-phone)
(contact,agent-phone)
(contact-phone, ? )
(phone,agent-phone)
(phone,agent-phone)
(listed-price,price)
(listed-price,price)
contact-phone => (agent-phone,0.7), (office-address,0.3)
Naive Bayes Learner [Domingos&Pazzani 97]
“Kent, WA” => (address,0.8), (name,0.2)
Whirl Learner [Cohen&Hirsh 98]
XML Learner
exploits hierarchical structure of XML data
Points:
1) Describe learners in general
2) Describe two example learners in detail
Note: must say clearly and briefly how each example learner works.

19. Training the Base Learners
Mediated schema
address price agent-phone description
location listed-price phone comments
Schema of realestate.com
Name Learner
<location> Miami, FL </>
<listed-price> $250,000</>
<phone> (305) </>
<comments> Fantastic house </>
(location, address)
(listed-price, price)
(phone, agent-phone)
...
realestate.com
Points:
1) Introduce our approach.
2) We do not manually map the schemas of all sources to mediated schema.
The goal is to manually mark up only a few sources, and be able to learn from the marked up sources to successfully propose mappings for subsequent sources.
3) Once the markup is done, there are many different types of information to learn from.
IT’S IMPORTANT TO SAY HERE THAT TRAINING = GLEANING KNOWLEDGE FROM THE DATA
<location> Boston, MA </>
<listed-price> $110,000</>
<phone> (617) </>
<comments> Great location </>
Naive Bayes Learner
(“Miami, FL”, address)
(“$ 250,000”, price)
(“(305) ”, agent-phone)
...

20. Entity Recognizers
Use pre-programmed knowledge to identify specific types of entities
date, time, city, zip code, name, etc
house-area (30 X 70, 500 sq. ft.)
county-name recognizer
Recognizers often have nice characteristics
easy to construct
many off-the-self research & commercial products
applicable across many domains
help with special cases that are hard to learn

21. Meta-Learner: Stacking
Training of meta-learner produces a weight for every pair of:
(base-learner, mediated-schema element)
weight(Name-Learner,address) = 0.1
weight(Naive-Bayes,address) = 0.9
Combining predictions of meta-learner:
computes weighted sum of base-learner confidence scores
Name Learner
Naive Bayes
(address,0.6)
(address,0.8)
<area>Seattle, WA</>
Meta-Learner
(address, 0.6* *0.9 = 0.78)

22. Training the Meta-Learner
For address
Extracted XML Instances
Name Learner
Naive Bayes
True Predictions
<location> Miami, FL</>
<listed-price> $250,000</>
<area> Seattle, WA </>
<house-addr>Kent, WA</>
<num-baths>3</>
...
Least-Squares Linear Regression
Weight(Name-Learner,address) = 0.1
Weight(Naive-Bayes,address) = 0.9

23. Applying the Learners Schema of homes.com Mediated schema
area day-phone extra-info
address price agent-phone description
Name Learner
Naive Bayes
<area>Seattle, WA</>
<area>Kent, WA</>
<area>Austin, TX</>
Meta-Learner
(address,0.8), (description,0.2)
(address,0.6), (description,0.4)
(address,0.7), (description,0.3)
Name Learner
Naive Bayes
Meta-Learner
(address,0.7), (description,0.3)
<day-phone>(278) </>
<day-phone>(617) </>
<day-phone>(512) </>
(agent-phone,0.9), (description,0.1)
Points:
1) Explain how the example learners are applied to match “area” with “address”.
2) Note that in general an arbitrary number of learners can be plugged into the matching process. If we have a new learner that has been trained, we can simply plug it in. This show how easy it is to add a new learner in our approach.
<extra-info>Beautiful yard</>
<extra-info>Great beach</>
<extra-info>Close to Seattle</>
(description,0.8), (address,0.2)

24. The Constraint Handler
Extends learning to incorporate constraints
hard constraints
a = address & b = address a = b
a = house-id a is a key
a = agent-info & b = agent-name b is nested in a
soft constraints
a = agent-phone & b = agent-name a & b are usually close to each other
user feedback = hard or soft constraints
Details in [Doan et. al., SIGMOD 2001]

25. The Current LSD System Training Phase Matching Phase Mediated schema
Source schemas
Domain
Constraints
Data listings
User Feedback
Constraint Handler
Base-Learner1
Base-Learnerk
Meta-Learner
Mappings

26. Outline Overview of structure mapping
Data integration and source mappings
LSD architecture and details
Experimental results
Current work.

27. Empirical Evaluation Four domains For each domain
Real Estate I & II, Course Offerings, Faculty Listings
For each domain
create mediated DTD & domain constraints
choose five sources
extract & convert data listings into XML (faithful to schema!)
mediated DTDs: elements, source DTDs:
Ten runs for each experiment - in each run:
manually provide 1-1 mappings for 3 sources
ask LSD to propose mappings for remaining 2 sources
accuracy = % of 1-1 mappings correctly identified

28. Matching Accuracy LSD’s accuracy: 71 - 92%
Average Matching Acccuracy (%)
LSD’s accuracy: %
Best single base learner: %
+ Meta-learner: %
+ Constraint handler: %
+ XML learner: %

29. Sensitivity to Amount of Available Data
Average matching accuracy (%)
Number of data listings per source (Real Estate I)

30. Contribution of Schema vs. Data
Average matching accuracy (%)
LSD with only schema info.
LSD with only data info.
Complete LSD
More experiments in the paper [Doan et. al. 01]

31. Reasons for Incorrect Matching
Unfamiliarity
suburb
solution: add a suburb-name recognizer
Insufficient information
correctly identified general type, failed to pinpoint exact type
<agent-name>Richard Smith</> <phone> (206) </>
solution: add a proximity learner
Subjectivity
house-style = description?

32. Outline Overview of structure mapping
Data integration and source mappings
LSD architecture and details
Experimental results
Current work.

33. Moving Up the Expressiveness Ladder
Schemas are very simple ontologies.
More expressive power = More domain constraints.
Mappings become more complex, but constraints provide more to learn from.
Non 1-1 mappings:
F1(A1,…,Am) = F2(B1,…,Bm)
Ontologies (of various flavors):
Class hierarchy (I.e., containment on unary relations)
Relationships between objects
Constraints on relationships

34. Finding Non 1-1 Mappings Current work
Given two schemas, find
1-many mappings: address = concat(city,state)
many-1: half-baths + full-baths = num-baths
many-many: concat(addr-line1,addr-line2) = concat(street,city,state)
1-many mappings
expressed as query
value correspondence expression: room-rate = rate * (1 + tax-rate)
relationship: state of tax-rate = state of hotel that has rate
special case: 1-many mappings between two relational tables
Flat schemas so that the set of operators under consideration can be simplified
Mediated schema
Source schema
address description num-baths
city state comments half-baths full-baths

35. Brute-Force Solution Define a set of operators
concat, +, -, *, /, etc
For each set of mediated-schema columns
enumerate all possible mappings
evaluate & return best mapping
Mediated-schema columns
Source-schema columns
compute similarity
using all base learners m1
m1, m2, ..., mk

36. Search-Based Solution
States = columns
goal state: mediated-schema column
initial states: all source-schema columns
use 1-1 matching to reduce the set of initial states
Operators: concat, +, -, *, /, etc
Column-similarity:
use all base learners + recognizers

37. Multi-Strategy Search
Use a set of expert modules: L1, L2, ..., Ln
Each module
applies to only certain types of mediated-schema column
searches a small subspace
uses a cheap similarity measure to compare columns
Example
L1: text; concat; TF/IDF
L2: numeric; +, -, *, /; [Ho et. al. 2000]
L3: address; concat; Naive Bayes
Search techniques
beam search as default
specialized, do not have to materialize columns

38. Multi-Strategy Search (cont’d)
Apply all applicable expert modules
L1: m11, m12, m13, ..., m1x
L2: m21, m22, m23, ..., m2y
L3: m31, m32, m33, ..., m3z
Combine modules’ predictions & select the best one
compute similarity
using all base learners
m11, m12,
m21, m22,
m31,m32
m11

39. Related Work ? Recognizers + Schema + 1-1 Matching
Single Learner Matching
TRANSCM [Milo&Zohar98]
ARTEMIS [Castano&Antonellis99]
[Palopoli et. al. 98]
CUPID [Madhavan et. al. 01]
SEMINT [Li&Clifton94]
ILA [Perkowitz&Etzioni95]
DELTA [Clifton et. al. 97]
Hybrid Matching
DELTA [Clifton et. al. 97]
Multi-Strategy Learning
Learners + Recognizers
Schema + Data
1-1 + non 1-1 Matching
Schema + Data
1-1 + non 1-1 Matching
Sophisticated Data-Driven User Interaction
LSD [Doan et. al. 2000, 2001]
CLIO [Miller et. al. 00],[Yan et. al. 01] ?

40. Summary LSD: Future work and issues to ponder:
uses multi-strategy learning to semi-automatically generate semantic mappings.
LSD is extensible and incorporates domain and user knowledge, and previous techniques.
Experimental results show the approach is very promising.
Future work and issues to ponder:
Accommodating more expressive languages: ontologies
Reuse of learned concepts from related domains.
Semantics?
Data management is a fertile area for Machine Learning research!

41. Backup Slides

42. Mapping Maintenance Ten months later ...
Mediated-schema M
Source-schema S m1 m2 m3
Ten months later ...
are the mappings still correct?
Mediated-schema M’
Source-schema S’ m1 m2 m3

43. Information Extraction from Text
Extract data fragments from text documents
date, location, & victim’s name from a news article
Intensive research on free-text documents
Many documents do have substantial structure
XML pages, name card, tables, list
Each such document = a data source
structure forms a schema
only one data value per schema element
“real” data source has many data values per schema element
Ongoing research in the IE community

44. Contribution of Each Component
Average Matching Acccuracy (%)
Without Name Learner
Without Naive Bayes
Without Whirl Learner
Without Constraint Handler
The complete LSD system

45. Exploiting Hierarchical Structure
Existing learners flatten out all structures
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
<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>

46. Domain Constraints Impose semantic regularities on sources Examples
verified using schema or data
Examples
a = address & b = address a = b
a = house-id a is a key
a = agent-info & b = agent-name b is nested in a
Can be specified up front
when creating mediated schema
independent of any actual source schema

47. The Constraint Handler
Predictions from Meta-Learner
Domain Constraints
a = address & b = adderss a = b
area: (address,0.7), (description,0.3)
contact-phone: (agent-phone,0.9), (description,0.1)
extra-info: (address,0.6), (description,0.4)
0.3
0.1
0.4
0.012
area: address
contact-phone: agent-phone
extra-info: address
0.7
0.9
0.6
0.378
area: address
contact-phone: agent-phone
extra-info: description
0.7
0.9
0.4
0.252
Can specify arbitrary constraints
User feedback = domain constraint
ad-id = house-id
Extended to handle domain heuristics
a = agent-phone & b = agent-name a & b are usually close to each other