Learning to Match Ontologies on the Semantic Web AnHai Doan Jayant Madhavan Robin Dhamankar Pedro Domingos Alon Halevy

Glue Identifies Mappings between websites Uses Machine Learning Uses Common Sense Knowledge Domain Constraints

Motivation Data comes from Different Ontologies Answers come from multiple web pages Manual: very tedious, error prone, not very scalable

Outline Overview of GLUE GLUE Architecture Case Studies CGLUE Case Studies Conclusion Assessment

Overview Assumes 2 Ontologies 1-1 Matching Similarity between two Concepts Computing Joint Distribution P(A,B), P(A, ~B), P(~A,B), P(~A,~B) Machine Learning Multistrategy Learning Exploiting Domain Constraints Data Instances

Overview Relaxation Labeler Similarity Estimator Meta Learner M L1L1 LkLk Taxonomy 0 1 Taxonomy 0 2 Joint Distributions Similarity function Similarity Matrix Common knowledge Domain constraints Mappings for Taxonomies …………

Distribution Estimator Meta Learner M Base Learner L 1 ………… Base Learner L k Taxonomy 0 1 Taxonomy 0 2 Joint Distributions

Distribution Estimator R DCA F E t1,t2 t3,t4 t5 t6,t7 t1,t2,t3,t4 t5,t6,t7 Trained Learner L

Distribution Estimator G H B JI s2,s3 s4 s5,s6 s1,s2,s3,s4 s5,s6 L s1

Distribution Estimator s1,s3 s5s6 s2,s4

Multistrategy Learning Base Learners Content Learner Frequency Naïve Bayes Name Learner Full Name Specific and Descriptive Element MetaLearner

Combines the base learners Gives learner weight User Input

Joint Distributions Similarity function Similarity Estimator Similarity Matrix Similarity Estimator

Applies Function From User Jaccard-sim Outputs a matrix between concepts

Where are we? Find Similarities Compute Similarities Satisfy Constraints

Relaxation Labeler Similarity Matrix Common knowledge Domain constraints Mappings for Taxonomies

Constraints Domain-Independent General Knowledge Domain-Dependent Interaction between two nodes Model each as a feature f()

Domain Independent

Relaxation Labeler Searches for best mapping given constraints Labels are influenced by it “neighborhood” Performs local optimization

Local Optimization 1. Assigns initial labels 2. Performs Optimization 3. Uses a formula to change a label 4. Repeat 2-3

Local Optimization Node in taxonomy O 1 Label in taxonomy O 2 Everything we know Other label assignments to all Nodes besides X

Local Optimization

Where are we? Relaxation Labeler Similarity Estimator Meta Learner M L1L1 LkLk Taxonomy 0 1 Taxonomy 0 2 Joint Distributions Similarity function Similarity Matrix Common knowledge Domain constraints Mappings for Taxonomies …………

Case Study University Catalogs Business Profiles For Each one Entire set of data instances Cleaned it up

Results

Improvements Insufficient Training Data Local Optimization Additional Base Learners Ambiguous Best Match

CGLUE

Beam Search Uses structure and data No relaxation labeling (no constraints)

CGLUE Case Study

Improvements Incorporate Domain Constraints Object Identification

Conclusion Semantic Similarity Multistategy Learning Relaxation Labeling CGLUE

Assessment Data Instances Additional Sites? CGLUE Future Work