Contextual ontology alignment may 2011

Contextual Ontology Alignment of LOD with an Upper Ontology: A Case Study with PROTON Prateek Jain, Peter Z. Yeh, Kumal Verma, Reymond Vasques, Mariana Damova, Pascal Hitzler and Amit Sheth ESWC’2011

Outline
Introduction
Problem and Conceptual Solution
Approach
Evaluation
Related Work
Conclusion

Linking Open Data (LOD)
Providers creating links across datasets
203 datasets in LOD cover a wide spectrum of subject domains – biomedical, science, geographic, generic knowledge, entertainment, government.
The linkage of these ever growing data sources takes place at the instance-level

Linking Open Data (LOD)
Without schema-level linkages the LOD cloud will not have semantic enough information to enable reasoning-based applications
Question Answering
Agent-based information brokering
and the like

Managing Linking Open Data (LOD)
FactForge ( http://factforge.net )
a reason-able view of the web of data
the biggest and most heterogeneous body of factual knowledge on which inference is performed
8 datasets from the LOD cloud (DBPedia, Freebase, UMBEL, CIA World Factbook , MusicBrainz, Wordnet, Lingvoj, Geonames)
an overall of 1, 4 million loaded statements, 2,2 million stored statements (indexed), 9,8 million distinct retrievable statements
manually developed schema-level alignment to efficiently query across multiple datasets

Outline
Introduction
Approach
Evaluation
Related Work
Conclusion

The Problem
Manual creation of schema-level mappings across LOD ontologies is not a viable solution
Size of the LOD
Rate at which it is growing
An automated solution is needed so that applications such as FactForge can effectively scale and keep up with the size of LOD.

Solution : BLOOM+ - Boothstraping-based LOD Ontology Matching
Extended BLOOMS
- uses more sophisticated metric to determine which classes between two ontologies to align
- considers contextual information to further support or reject an alignment

Outline
Introduction
Approach
Evaluation
Related Work
Conclusion

Access to FactForge
Forest Interface
- keyword search (for molecules in the RDF graph)
- auto-suggestion of URI
- SPARQL queries using facts from different datasets
( formulating SPARQL queries requires in depth knowledge about the datasets and the schemata in FactForge )
- exploration

Knowledge Requirements
A knowledge source:
- organized as a class hierarchy
links between classes capture sub and super class relationships
- covering a wide range of concepts and domains
widely applicable, given the wide range of domains covered by the LOD cloud
Wikipedia category hierarchy is used by BLOOMS+
taxonomy of over 10 million categories across many domains.

Approach
BLOOMS+ aligns two ontologies with the steps:
- Use Wikipedia to construct a set of category trees for each class in the source and target ontologies
- Determine which classes to align
- a sophisticated measure to compute similarity between source and target classes
- contextual similarity between the classes to support or reject an alignment
- Align classes with high similarity based on the class and contextual similarity

Construct BLOOMS+ Forest
The root of the tree is s i
The immediate children of s i are all categoies that s i belongs to
Each subsequent level includes all unique, direct super categories of the categories at the current level
BLOOMS+ imposes a limit on the depth of the tree 4 as it has been empirically proven that depths beyond 4 typically include very general categories not useful for alignment.

Compute Class Similarity
Comparison between each class C from the source ontology with each class D from the target ontology
each T i  F c with each T j  F d
For each T i its overlap with T j is determined
Issues:
- common nodes appear often, result in false alignments
- large tree can be unfairly penalized because it must have more nodes in common with another three for a high similarity score

Compute Class Similarity
n  Ti  Tj the common nodes
d(n) is the depth of the common node in Ti
The exponentiation of the inverse depth of a common node gives less importance to the node if it is generic
The log of the tree size avoids bias against large trees
Range from 0.0 to 1.0
Overlap ( Ti , Tj ) = log  n  Ti  Tj (1+e d(n)-1 -1 ) log2 | Ti | To cope with the issues:

Compute Contextual Similarity
A good source of contextual information is the superclass of C and D from their respective ontologies.
If superclasses agree, then alignment is supported, if not alignment is penalized.
The method:
- for each pair wise class comparison ( C , D ) all superclasses of C and D are retrieved up to a specified level of 2.
- the two sets of superclasses N(C) and N(D) are the neighbourhoods of C and D .
- for each tree pair ( Ti , Tj ) between C and D , the number of superclasses in N(C) and N(D) that are supported by Ti and Tj are determined.
c  N(C) is supported by T i if
- the name of c matches a node in T i
- a Wikipedia article corresponding to c matches a node in Ti

Compute Contextual Similarity
The overall contextual similarity between C and D with respect to Ti and Tj is computed with a harmonic mean:
CSim(Ti, Tj) = 2R C R D R C + R D
R C and R D are the fractions of superclasses in N(C) and N(D) supported by Ti and Tj
This harmonic mean helps emphasizing the superclass neighbourhoods that are not well supported thus lowering the overall contextual similarity.

Compute Overall Similarity
The overall similarity between classes C and D w.r.t. BLOOMS+ trees Ti and Tj is computed by taking the weighted average of the class and contextual similarity:
O(Ti, Tj) =  Overlap(Ti , Tj) +  CSim(Ti , Tj)
2
and  are weights for the concept and contextual similarity
Both are defaulted to 1.0 to give an equal importance to the two parameters.
Selected is the tree pair (Ti, Tj)  F C  F D with the highest overall similarity score and if the score is greater than the alignment threshold H A . Then a link between C and D will be established.

Determining of the Type of Link
If O(Ti, Tj) = O(Tj, Ti) then C owl:equivalentClass D
If O(Ti, Tj) < O(Tj, Ti) then C rdfs:subClassOf D
Otherwise, D rdfs:subClassOf C

Outline
Introduction
Approach
Evaluation
Related Work
Conclusion

Evaluation
Criteria:
BLOOMS+ can outperform state-of-the-art solutions on the task of aligning LOD ontologies
BLOOMS+ performs well because it accounts for the critical factors when computing the similarity between two classes
The importance of common nodes between the BLOOMS+ trees for the two classes
Bias against large trees
The performance of BLOOMS+ can be further improved by using contextual information

Dataset – the Gold Standard
Schema-level mappings from 3 LOD datasets to PROTON
PROTON – an upper-level ontology with over 300 classes and 100 properties designed to support semantic annotation, indexing and search
DBpedia – the RDF version of Wikipedia, of 259 classes (Event to Protein)
Freebase – a knowledge base of structured data collected from multiple sources
Geonames – a geographic dataset with over 6 million locations of interest, which are classified into 11 different classes
These mappings have been systematically created by knowledge engineers at Ontotext for FactForge, which enables SPARQL query over the LOD cloud
544 mappings
373 PROTON-DBpedia
150 PROTON-Freebase
12 PROTON-Geonames

Dataset – the Gold Standard
Advantages
The mappings were created by an independent source for a real world use case
The mappings were created by knowledge engineers through a systematic process, hence their high quality
The mappings cover a diverse set of ontologies

Experimental Setup
(a) BLOOMS+ can outperform state-of-the-art solutions on the task of aligning LOD ontologies
- Precision and Recall of the mappings from the LOD to PROTON
BLOOMS+ was applied to each PROTON-LOD ontology pair with alignment threshold of 0.85, because empirically this threshold produces the best F-measure
Then the results were compared with the Gold Standard
Precision : the number of correct mappings over the total number of mappings generated by BLOOMS+
Recall : the number of correct mappings over all mappings in the gold standard

Experimental Setup
(a) BLOOMS+ can outperform state-of-the-art solutions on the task of aligning LOD ontologies
- Precision and Recall of the mappings from the LOD to PROTON
Comparison with the results of
BLOOMS
S-Match (three matching algorithms – basic, minimal, structure preserving)
AROMA (association rule mining paradigm to discover relationaships)
The best alignment threshold for BLOOMS is 0.6

Experimental Setup
(b) BLOOMS+ performs well because it accounts for the critical factors when computing the similarity between two classes
The importance of common nodes between the BLOOMS+ trees for the two classes
Bias against large trees
(c) The performance of BLOOMS+ can be further improved by using contextual information
BLOOMS+ without contextual information vs. BLOOMS
(b) measure used to compute the similarity between two classes
BLOOMS+ without contextual information vs. BLOOMS+
(c) use of contextual information
Alignment threshold to 0.85

Experimental Results Linked Open Data and Proton Schema Ontology Alignment DB-PRO GEO-PRO FB-PRO Overall System Rec Prec F Rec Prec F Rec Prec F Rec Prec F AROMA 0.19 0.59 0.28 0.04 8/1000 0.01 0.31 0.49 0.38 0.22 0.37 0.28 S-Match-M 0.26 0.05 0.08 0.04 6/1000 0.01 0.2 0.05 0.08 0.23 0.05 0.08 S-Match-C 0.33 3/1000 6/1000 0.04 0.009 0.01 0.3 0.4 0.34 0.31 4/1000 0.007 BLOOMS 0.48 0.19 0.27 0.04 6/1000 0.01 0.28 0.32 0.3 0.42 0.19 0.26 BLOOMS+ no context 0.77 0.59 0.67 0.04 5/1000 0.01 0.48 0.65 0.55 0.66 0.45 0.54 BLOOMS+ 0.73 0.90 0.81 0.04 5/1000 0.01 0.49 0.59 0.54 0.63 0.55 0.59

Outline
Introduction
Approach
Evaluation
Related Work
Conclusion

Related Work Euzenat, J. & al. Matching ontologies for context, a Tech Report of NEON project from 2007 - rely on background knowledge from online ontologies - their process relies on identification of contextual relationship using the relationships encoded in the ontologies Ontology matching surveys (Euzenat and Shvaiko, 2007) emphasize that systems typically utilize a structured source of information (dictionaries or upper level ontologies) Wikipedia categorization has been utilized for creating and restructuring taxonomies Identification and creation of links between LOD cloud datasets ontology schema matching used to improve instance coreference resolution UMBEL – a unified reference point to LOD schemas

Outline
Introduction
Approach
Evaluation
Related Work
Conclusion

Conclusion
A solution for automated ontology matching – BLOOMS+
Evaluated on a real world Gold Standard mapping LOD ontologies to PROTON created for FactForge
Compared the performance of BLOOMS+ to other state of the art systems
BLOOMS+ is the only system utilizing the contextual information present in the ontology and Wikipedia category hierarchy for OM

Future Work
Use BLOOMS+ for LOD querying
Use other techniques to improve BLOOMS+
Test BLOOMS+ with other knowledge sources
Incorporate additional contextual information
Use BLOOMS+ in data integration scenarios

Thank you for your attention!
Questions?

Contextual ontology alignment may 2011

by Mariana Damova

Accessibility

Categories

Upload Details

Usage Rights

2 Embeds 84

Statistics