1. Semantic Mappings for Data Mediation
Jayant Madhavan
University of Washington
Joint work with AnHai Doan, Pedro Domingos, and Alon Halevy
Good afternoon.
Jayant Madhavan, Semantic Mappings for Data Mediation
Joint work with AnHai Doan and professors Pedro Domingos and Alon Halevy
I shall start with some motivation for the problem, and then go on to sketch a brief outline of two systems that we have built here at UW.

2. Find houses with 2 bedrooms priced under 300K Charlie comes to town
homes.com
realestate.com
homeseekers.com
I shall start with a motivating example.
Charlie comes to Seattle… decides that he needs to buy a house – a 2 bedroom one priced in the 300K range
Being a man of the information age, he decides to take the help web-based agencies
Every week he looks up listings at sites such as realestate.com
Frustration soon creeps in… he’s having to go to easy site and do the search all over
26th Febraury, 2002
Affiliates Meeting

3. Find houses with 2 bedrooms
Data Integration
Find houses with 2 bedrooms
priced under 300K
mediated schema
source schema 1
source schema 2
source schema 3
Instead Charlie decided to use a data integration system… one that integrates, one that integrates sources for real estate listings.
Here’s how it works… each site publishes its listings in its source schema.
The system has a mediated schema – like a unified view.
Charlie’s queries are posed over the mediated schema, and the system translates them to queries over the different sources, and the results are later combined.
The green arrows enable this translation. They are what we shall call mappings.
They identify corresponding data entities in the mediated schema and the source schemas.
Our general approach for data integration will be to manually provide manual mappings from a few 2/3 data sources to the mediated schema, and the rest will be learnt automatically. So if 100 => do 3, other 97 done automatically.
wrapper
wrapper
wrapper
realestate.com
homeseekers.com
homes.com
26th Febraury, 2002
Affiliates Meeting

4. Semantic Mappings between Schemas
Mediated schema
address agent-name agent-city agent-state
1-1 mapping
complex mapping
homes.com
area contact-name contact-address
Denver, CO Laura Smith Boulder, CO
Oakland, CA Jean Brown Davis, CA
Here’s a closer look at an example mapping.
1-1 mappings and more complex mappings
Our focus on 1-1 mappings
26th Febraury, 2002
Affiliates Meeting

5. Why Schema Matching is Important
Enterprise 1
Data integration
Data integration
Data translation
Data warehousing
E-commerce
World-Wide Web
Ontology
Matching
Mapping are relevant in many-many application areas.
Other examples – data translation, migrating data from legacy database to a new format.
E-commerce – Mediation between two services that use different XML formats.
Ontology mapping – Ontologies are rich schemas that have been proposed recently as a means of tagging data in order to enable automatic interpretations by soft bots.
Mapping between multiple onto is a necessacity it the vision of automatic reasoning over the web is to be a reality.
In short, any applications
Knowledge
Base 2
Information
agent
Enterprise 2
Home users
Knowledge Base 1
Application has more than one schema need for schema matching!
26th Febraury, 2002
Affiliates Meeting

6. Why Schema Matching is Difficult
No access to exact semantics of concepts
Semantics not documented in sufficient details
Schemas not adequately expressive to capture semantics
Must rely on clues in schema & data
Using names, structures, types, data values, etc.
Such clues can be unreliable
Synonyms: Different names => same entity:
area & address => location
Homonyms: Same names => different entities:
area => location or square-feet
Done manually by domain experts
Expensive and time consuming
Obtaining these mappings would have been easy if were aware of the exact semantics of each and every dat entity or concept.
Not the case, schemas not well documented, or schemas might not reveal the complete semantics.
Must rely of heuristics that use many different clues available in the schemas and the data, such as using the different names of entities, their data types, sample values, etc.
Among other things these clues are often unreliable – synonymy and homonymy
For most part done manually by domain experts, and the process is usually expensive and time consuming.
26th Febraury, 2002
Affiliates Meeting

7. Previous work Mostly ad-hoc heuristics Graph matching
Name matchers
Data types
Sample domain values
Graph matching
Schemas are labeled graphs
No single heuristic works across scenarios
Systems are fragile and need a lot of tuning
Most prior work has been in the form of ad-hoc heuristics,
Name matchers that look for synonyms, common suffixes/prefixes, etc.,
Some that use the data types and others that use sample domain values
There is bunch of work on graph matching based approaches that interpret schemas as labeled graphs
Some works combine different approaches also in some ad-hoc fashion
Heuristics successful to various extents in different scenarios.
No single one works
Systems are fragile and need a lot of tuning.
26th Febraury, 2002
Affiliates Meeting

8. How do we go about it? Make extensive use of data instances
Incorporate multiple heuristics
Base learners that implement individual heuristics
Machine Learning
Multi-strategy learning to combine base learners
Extensible framework
Easy to add new heuristics/learners
Generic and domain specific constraints
Robust solution with high accuracy
We’ve built matching solutions for two scenarios.
Basic underlying characteristics
- Extensive use of data instances, as opposed to just schema information.
- Incorporate multiple heuristics. Each of these in our implementation is a learning algorithm.
- Predictions of base learners combined by multi-strategy.
- Easy to add new heuristics/learners
- Easy to add a large set of expressive constraints.
=> Robust solution with accuracy
26th Febraury, 2002
Affiliates Meeting

9. If “office” occurs in the name => office-phone
Multiple hypotheses
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) (305) Fantastic house
Boston, MA $320K Mike Doan (617) (617) Great location
If “fantastic” & “great” occur frequently in data instances
=> description
Examples of how multiple heuristics work
Two sources of information – data contents and concept names.
Content matcher example.
Name matcher, note that CM will not work here.
Combine hypotheses.
If “office” occurs in the name => office-phone
Content matcher
Name matcher
26th Febraury, 2002
Affiliates Meeting

10. Base Learners Mediated schema realestate.com
address price agent-name agent-phone office-phone description
realestate.com
location price contact-name contact-phone office comments
(“location”, address)
(“price”, price)
(“contact name”, agent-name)
(“contact phone”, agent-phone)
(“office”, office-phone)
(“comments”, description)
Name Learner
Miami, FL $250K James Smith (305) (305) Fantastic house
Boston, MA $320K Mike Doan (617) (617) Great location
Content Learner
(“Miami, FL”, address)
(“$250K”, price)
(“James Smith”, agent-name)
(“(305) ”, agent-phone)
(“(305) ”, office-phone)
(“Fantastic house”, description)
(“Boston,MA”, address)
Here’s how we train our base learners –
As a reminder, we manually map our first few data sources to the mediated schema.
We extract training data for our Content Learner by looking at data instances and assigning them their corresponding label in the mediated schema.
Same too for the name matcher, but here just the element names.
Results in learnt hypotheses like earlier
26th Febraury, 2002
Affiliates Meeting

11. Learning Source Descriptions (LSD) [SIGMOD’01]
Training Phase
Mediated schema
Source schemas
Matching Phase
Meta-Learner
Base-Learner Base-Learnerk
Predictions for data instances
Base-Learner1
Base-Learnerk
Training data
for base learners
Hypothesis1
Hypothesisk
Prediction Combiner
Predictions for elements
A look at the complete LSD solution for learning semantic mappings for data integration
- Trained by manual mappings
- Training data extracted for base learners
- Meta learner is trained using stacking to obtain weights that indicate the importance of individual learners for make each prediction in the mediated schema
- Matching phase
- Given a new data source, we apply each individual base learner, and then combine the predictions using the meta-learner
- First combined at the data instance level and then at the element or concept level
- Apply a set of constraints that can be specified by the user to prune away candidate mappings.
- Finally produce mappings, can be corrected by user feedback and fed back to the system as a constraint.
Constraint Handler
Mappings
Domain
constraints
Weights for
Base Learners
Meta-Learner
26th Febraury, 2002
Affiliates Meeting

12. LSD’s performance LSD’s accuracy: 71 - 92%
Avg. Matching Accuracy (%)
LSD’s accuracy: %
Best single base learner: %
+ Meta-learner: %
+ Constraint handler: %
Complete LSD system: %
26th Febraury, 2002
Affiliates Meeting

13. Matching Ontologies of Concepts
CS Dept Australia
CS Dept U.S.
Undergrad
Courses
Grad
Courses
Courses
People
Staff
Faculty
Staff
Academic Staff
Technical Staff
Senior
Lecturer
Assistant
Professor
Associate
Professor
Professor
Lecturer
Professor
Different problem – ontology matching
- Ontology is a conceptualization of a domain that identifies concepts in the domain, and their inter-relationships.
- Inheritance tree (taxonomy) that identifies gens/specs There are data instances in at the leaves.
- For each concept, find the most similar concept
Each ontology has an inheritance tree (taxonomy) and data instances at the leaves.
For each concept find most similar concept in the other ontology.
26th Febraury, 2002
Affiliates Meeting

14. The Glue System [WWW’2002] No manually performed mappings
Automatically collect training data for base learners.
Similarity measures computed from the joint probability distribution of concepts
A random data instance can belong to both, either, neither concepts – P(A,B), P(A,B’), P(A’,B), P(A’,B’).
General framework for incorporating constraints
Extension of relaxation labeling.
26th Febraury, 2002
Affiliates Meeting

15. The Glue System Relaxation Labeling Similarity Estimator Meta Learner
Mappings for O1 , Mappings for O2
Relaxation Labeling
Similarity Matrix
Common Knowledge &
Domain Constraints
Similarity Estimator
Joint Probability Distribution P(A,B), P(A’, B)…
Similarity Function
Meta Learner
Distribution
Estimator
Base Learner
Base Learner
Taxonomy O1
(tree structure + data instances)
Taxonomy O2
(tree structure + data instances)
26th Febraury, 2002
Affiliates Meeting

16. Glue’s performance
26th Febraury, 2002
Affiliates Meeting

17. Conclusion and Future Work
LSD and Glue perform well
Combine predictions of different base learners
Incorporate constraints
Robust solution that results in good accuracy
Future Work
Representation mapping system
Incorporates various heuristics
Can perform complex mappings
Can learn with experience.
Reasoning about mappings
Does a mapping enable answering of queries posed on other schema?
Is one mapping implied by another? Is a mapping minimal?
Can mappings be composed?
26th Febraury, 2002
Affiliates Meeting