Schema Mapping of Formbased Web Interfaces

1
Schema Mapping of Form-based Web Interfaces

• Aravind Kalavagattu
• Raju Balakrishnan
• Garrett Wolf
• Yochan Information Integration Group

2
Motivation

• Deep Web Problem on the WWW
• Tons of data reside in the hidden databases
• Many of current web search engines fail to crawl
  this valuable data
• Data Integrators
• A data integration application works by
  integrating data from several online databases
• They all need Schema Mapping as their first and
  foremost step to query/crawl the databases!
• Many sites provide only a restricted query access
  through form-based interfaces.

3
Problem Statement

• Formally, the web-interface is a schema F with n
  form-fields f1, f2, f3fn.

• Given two html form-schemas G and H, the problem
  of semantic schema mapping boils down to mapping
  a subset of fields in G, i.e., (g1 g2 g3 gm)
  with the subset of fields in H, i.e., (h1 h2
  h3 hn) such that each
• gi corresponds to hi.
• For example, in the above picture, taking G as,
  cars.com, and H as, Autoads.com, we have the
  mappings G.Make H.Make, G.Model H.Model,
  G.MaxPrice H.Price, G.SearchWithin
  H.Distance, G.YourZip H.ZipCode.

4
Steps Involved

• Feature Identification
• Distance/Similarity Metrics
• Methodology
• Feature Weights Learning
• Clustering Approach
• SVM based Classification
• Empirical Evaluation
• Results and Discussion

5
Feature Identification

• Name
• Actual name of the html input element
• Description
• The description commonly located to the left or
  above the field when viewing a rendered HTML page
• Values List
• Applies to selection lists and radio buttons
• HTML Control Type
• Type selection list, text box, check box, radio
  button
• Data Type
• Data type of the values

6
Similarity Metrics

• Name Similarity (NS)
• Two types
• Jaccard Coefficient
• Semantic similarity
• Distance between the common hypernym of the two
  words
• Example Maharashtra Arizona are similar
  now.
• Description Similarity (DS)
• Same as above
• Value List Similarity (VLS)
• Jaccard Coefficient
• Control Type Similarity (CTS)
• For differing types, we decide if one subsumes
  the other
• Data Type Similarity (DTS)
• Its in a manner similar to the above similarity

7
Steps Involved

• Feature Identification
• Distance/Similarity Metrics
• Methodology
• Feature Weights Learning
• Clustering Approach
• SVM based Classification
• Empirical Evaluation
• Results and Discussion

8
Learning Feature Weights

• The similarity between two form fields f and g,
  will be
• Sim(f,g) WT.SIMfgb
• Perceptron based gradient descent, with ideal
  kernel matrix element as target values to
  optimize weights
• Perceptron learning Equal number of positive
  and negative samples are used across domains,

9
Clustering Approach

• Similarity matrix S, where Sij weighted
  similarity between the form-fields i and j, we
  perform a agglomerative hierarchical clustering
  using the three different distance measure
• Single-link, Complete-link, Average-link

10
SVM Classification Approach

• Kernel gram matrix formed as weighted sum of
  individual similarity measures.
• The negative eign values are set to zero to make
  the kernel matrix positive semi-definite.
• N one-vs-all SVMs classifiers were trained to
  classify fields, where N is number of classes,
  with five
• The Trained SVM is tested Against the test data

11
Steps Involved

• Feature Identification
• Distance/Similarity Metrics
• Methodology
• Feature Weights Learning
• Clustering Approach
• SVM based Classification
• Empirical Evaluation
• Results and Discussion

12
Experiments

• Distance Metrics
• Bag of Words Similarity (Jaccard)
• name, description, values
• Semantic Similarity (WordNet)
• name, description
• Is-A Relationship Similarity (Attribute
  Hierarchy)
• data type
• Boolean Similarity (Exact Match Only)
• html type

• Data Set
• UIUC Information Integration Repository
• TEL-8 Query Interfaces data set
• 4 of 8 domains
• Airfares, Automobiles, Hotels, Car Rentals
• Features
• name
• description
• values
• data type
• html type

13
Which Distance Metric to Use?
14
How to Assign Feature Weights?
15
Can Weights be Used Across Domains?
What Affect do the Negative Similarities Have?

• Assigning negative similarity between fields in
  the same form improves accuracy when compared to
  zero similarity assignments.

16
Observations

• Clustering using Single-Link distance performs
  badly! Complete and Average perform equally
  good, but on the whole, Average-link tends to
  win!
• The perceptron approach to learn the feature
  weights does best compared to uniform/manual
  assignments!
• Clustering using Average link and
  perceptron-learned feature weights gave us high
  accuracy compared to the SVM based approach.
• Feature weights can be used across domains.

17
Thank You