DBpedia Release Evaluation: Difference between revisions

From DBpedia Mappings
Jump to navigationJump to search
No edit summary
No edit summary
Line 9: Line 9:
|+ Mapped triples (Mappings from 3.5.1 DBpedia release)
|+ Mapped triples (Mappings from 3.5.1 DBpedia release)
|-
|-
! rowspan="2" | Category
! rowspan="3" | Category
! rowspan="2" | Gold Triples
! rowspan="3" | Gold Triples
! colspan="2" | DBpedia Release
! colspan="2" | DBpedia Release
|-
|-

Revision as of 16:24, 14 September 2011

The Quality Assessment Framework (QAF) is developed to document the quality of the knowledge base and furthermore the progress of DBpedia's extraction framework. The main idea of the QAF is a comparison between a manually created best-case dataset (Gold Standard) and the output from DBpedia's ontology based extraction. The QAF estimates the precision of the extraction framework and the completeness (recall) of DBpedia compared to its source Wikipedia.

Sample Data / Gold Standard

For a significant evaluation, only potentially extractable triples are considered. Only if a triple arise from a mapped property it can be extracted. Here this triples are called mapped triples. The following table shows the number of mapped triples, the total number of triples in the Gold Standard and the percentage of mapped triples for each category. The categories result from different patterns in which the information is given in Wikipedia infoboxes. The number of cases differ in a high extent depending on the category. The results for the categories based on small numbers should be handled with care.

Mapped triples (Mappings from 3.5.1 DBpedia release)
Category Gold Triples DBpedia Release
DBpedia 3.5.1
Mapped Triples %
Total 1504 3221 46.7
Plain Property 514 893 57.6
Number-Unit 51 76 67.1
Coordinate 36 54 66.7
Interval 22 31 71.0
List 478 801 59.7
One-Property-Table 242 447 54.1
Multi-Poprty-Table 83 625 13.3
Open Property 13 139 9.4
Open Property Table 0 26 0.0
Internal Template 58 116 50.0
Merged Properties 7 13 53.8

Evaluation Results

The following two tables show the completeness (recall) and precision of the different release versions. To achieve comparability and to clean the results from the effect of more and better mappings, the mapping version of the 3.5.1 release is taken as constant.

Completeness / Recall (fixed mappings from 3.5.1)
Category DBpedia 3.5.1 DBpedia 3.6 DBpedia 3.7
Total 45,7% 60,2% 61,8%
Plain Property 80,5% 83,7% 86,0%
Number-Unit 68,6% 68,6% 66,7%
Coordinate 100,0% 100,0% 100,0%
Interval 72,7% 68,2% 72,7%
List 33,9% 75,7% 77,8%
One-Property-Table 5,4% 5,8% 5,8%
Multi-Poprty-Table 0,0% 0,0% 0,0%
Open Property 23,1% 23,1% 30,8%
Open Property Table na na na
Internal Template 8,6% 10,3% 10,3%
Merged Properties 57,1% 57,1% 71,4%
Precision (fixed mappings from 3.5.1)
Category DBpedia 3.5.1 DBpedia 3.6 DBpedia 3.7
Total 91,2% 92,3% 92,4%
Plain Property 96,3% 96,6% 97,4%
Number-Unit 85,4% 85,4% 85,0%
Coordinate 100,0% 100,0% 100,0%
Interval 100,0% 100,0% 88,9%
List 91,5% 92,8% 93,2%
One-Property-Table 32,5% 36,8% 34,1%
Multi-Poprty-Table na na na
Open Property 100,0% 100,0% 100,0%
Open Property Table na na na
Internal Template 83,3% 75,0% 75,0%
Merged Properties 80,0% 80,0% 100,0%