Mini-Ontology Generation from Canonicalized Tables

1
Mini-Ontology Generation from Canonicalized Tables

• Stephen Lynn
• Data Extraction Research Group
• Department of Computer Science
• Brigham Young University

Supported by the
2
TANGO Overview
TANGO Table ANalysis for Generating
Ontologies Project consists of the following
three components

1. Transform tables into a canonicalized form
2. Generate mini-ontologies
3. Merge into a growing ontology

3
Sample Input
Region and State Information Region and State Information Region and State Information Region and State Information
Location Population (2000) Latitude Longitude
Northeast 2,122,869
Delaware 817,376 45 -90
Maine 1,305,493 44 -93
Northwest 9,690,665
Oregon 3,559,547 45 -120
Washington 6,131,118 43 -120
Sample Output
4
Mini-Ontology GeneratOr (MOGO)

• Concept/Value Recognition
• Relationship Discovery
• Constraint Discovery

5
Concept/Value Recognition

• Lexical Clues
• Labels as data values
• Data value assignment
• Data Frame Clues
• Labels as data values
• Data value assignment
• Default
• Classifies any unclassified elements according to
  simple heuristic.

6
Relationship Discovery

• Dimension Tree Mappings
• Lexical Clues
• Generalization/Specialization
• Aggregation
• Data Frames
• Ontology Fragment Merge

7
Constraint Discovery

• Generalization/Specialization
• Computed Values
• Functional Relationships
• Optional Participation

Region and State Information Region and State Information Region and State Information Region and State Information
Location Population (2000) Latitude Longitude
Northeast 2,122,869
Delaware 817,376 45 -90
Maine 1,305,493 44 -93
Northwest 9,690,665
Oregon 3,559,547 45 -120
Washington 6,131,118 43 -120
8
Validation

• Concept/Value Recognition
• Correctly identified concepts
• Missed concepts
• False positives
• Data values assignment
• Relationship Discovery
• Valid relationship sets
• Invalid relationship sets
• Missed relationship sets
• Constraint Discovery
• Valid constraints
• Invalid constraints
• Missed constraints

Precision Recall F-measure
Concept Recognition 87 94 90
Relationship Discovery 73 81 77
Constraint Discovery 89 91 90
9
Concept Recognition

• What we counted
• Correct/Incorrect/Missing Concepts
• Correct/Incorrect/Missing Labels
• Data value assignments

10
Relationship Discovery

• What we counted
• Correct/incorrect/missing relationship sets
• Correct/incorrect/missing aggregations and
  generalization/specializations

11
Constraint Discovery

• What we counted
• Correct/Incorrect/Missing
• Generalization/Specialization constraints
• Computed value constraints
• Functional constraints
• Optional constraints

12
Concept Recognition

• Successes
• 98 of concepts identified
• Missing label identification
• 97 of values assigned to correct concept
• Common problems
• Finding an appropriate label
• Duplicate concepts

13
Relationship Discovery

• Recall of 92 for relationship sets
• Missing aggregations and generalizations/specializ
  ations
• Only found in label nesting

14
Constraint Discovery

• F-measure of 98 for functional relationship sets
• Poor computed value discovery
• Rows/Columns with totals

15
Conclusions

• Tool to generate mini-ontologies
• Assessment of accuracy of automatic generation

Precision Recall F-measure
Concept Recognition 87 94 90
Relationship Discovery 73 81 77
Constraint Discovery 89 91 90
16
Future Work

• Tool Enhancements
• Linguistic processing
• Data frame library
• Domain specific heuristics
• Alternate Uses
• Annotation for the Semantic Web