dringler / knowledgegraphanalysis

Code accompanying our paper "One Knowledge Graph to Rule them All? Analyzing the Differences between DBpedia, YAGO, Wikidata & co."

Knowledge Graph Analysis

Code accompanying our paper "One Knowledge Graph to Rule them All? Analyzing the Differences between DBpedia, YAGO, Wikidata & co."

Quantitative analysis of the following Knowledge Graphs (KGs):

• DBpedia (D)
• YAGO (Y)
• Wikidata (W)
• NELL (N)
• OpenCyc (O)

Approach:

• Get top 10 classes for each KG
• Calculation of class indegree and outdegree
• Get all instances for each class
• Calculation of minimum, average, median, and maximum indegree and outdegree for the instances of each class
• Create a combined list with all top 10 classes and equal classes in other KGs (e.g. with owl:sameAs properties)
• Calculate all degree values for the new classes as well
• Calculate the instance overlap of the classes using different string similarity measures

Instructions:

1. /LinkedInstances/*.py creates files with all linked instances between two KGs.
   • Input:
     • KG files containing instances and/or links to other instances.
   • Output:
     • Files containing the combined links between two KGs (e.g. DO_sameAs_union.nt for the links between DBpedia and OpenCyc) that are denoted as #o1.
     • Move those #o1 files to the /InstanceOverlap/owlSameAs/ folder.
2. /GetInstances/src/GetInstances.java creates files that contain all instances of a class including all English labels.
   • Input:
     • Array with class names for each KG.
     • Full KG or just the files containing the instances and labels.
   • Output:
     • Textfiles containing all instances with all English labels for each class in each KG.
     • Saved as <k_className>InstancesWithLabels.txt where k stands for the abbreviation of the KG (e.g. d_ActorInstancesWithLabels.txt for the actor instances in DBpedia). All those files are denoted as #o2.
     • Move these #o2 files to the /InstanceOverlap/InstanceLabels/ folder.
3. /InstanceOverlap/src/InstanceOverlapMain.java executes the following three steps for each class in the className array for calculating the estimated overlap:
   1. CountSameAs.java creates files with the linked instances of two classes by e.g. using the owl:sameAs property.
      • Input:
        • Class name.
        • #o1 files with the linked instances in the /InstanceOverlap/owlSameAs/ folder.
        • #o2 files with all English instance labels for the respective class and for each KG in the /InstanceOverlap/InstanceLabels/ folder.
      • Output:
        • Links between instances for each class1-class2 combination that is used as gold standard (there might be multiple classes that describe the same concept in a single KG, e.g. wordnet_actor_109765278 and wordnet_actor_109767197 in the YAGO KG). These files are saved as <className1_className2>.tsv in the /InstanceOverlap/owlSameAs/x2y/ folder (e.g. Actor_wordnet_actor_109765278.tsv in the d2y folder). These files are denoted as #o3.
   2. CountStringSimilarity.java creates files that contain all found links between two classes using the different string similarity measures (e.g. Jaro, Levenshtein) and different thresholds.
      • Input:
        • Class name.
        • #o2 files.
      • Output:
        • Links between the instances of two classes that are found using a specific similarity measure and threshold. The results are saved as <fromK_2_toK_fromClass_toClass_simMeasure_threshold>.tsv in the /InstanceOverlap/simMeasureResults/ folder (e.g. d2y_Actor_wordnet_actor_109765278_jaro_1.0.tsv). These files are denoted as #o4.
   3. EstimatedInstanceOverlap.java
      • Input:
        • Class name.
        • #o3 containing linked instances that is used as gold standard.
        • #o4 containing the instances that should be linked based on the respective similarity measure and threshold.
      • Output:
        • estimatedOverlap_<className_parameter_timestamp>.csv files in the /InstanceOverlap/estimatedOverlap/ folder containing instance counts, precision, recall, f-measure, estimatedOverlap, number of links, count of matching alignment, count of partial matching alignment, and true positives for each class1-class2 combination for each class and each KG combination (e.g. estimatedInstanceOverlap_Actor_wBlockingMax1000000_tokenBk4_2017_02_17_13_35_52.csv).

knowledgegraphanalysis's Issues

update WD to 2017-08

Excellent analysis and paper!
But WD 2016-08 is 14m old and current WD has grown from 17.5 entities to 37.2M. Would it be possible to update the WD analysis using current data?