Skip to content

Latest commit

 

History

History
165 lines (141 loc) · 6.9 KB

README.md

File metadata and controls

165 lines (141 loc) · 6.9 KB
Raw

DL-TC-Generator

Java Build DOI License

This maven project can generate a description logics gold standard to evaluate knowledge graph embeddings in terms of their ability to learn specific logical expressions. Multiple options exist to generate a gold standard. For the evaluation, this framework can be used.

Gold Standards

Overview

Version Name Location Based On Size Classes Other Parameters
v1 dbpedia /results/v1/dbpedia DBpedia 2021-09 50, 500, 5000 -
v1 synthetic_ontology /results/v1/synthetic_ontology -
1000
  • # classes: 760
  • # properties: 1355
  • Max triples per node: 11
  • AVG branching factor: 5
  • node factor: 10

DBpedia SPARQL Gold Standard

Given multiple SPARQL queries defined in the query directory, this project generates a gold standard that can be used for machine learning extensions.

The gold standard is intended to be used with a DBpedia embedding.

You can find the current version of the gold standard in the results directory.

Synthetic Gold Standard

The DBpedia SPARQL GS (see above) is a real gold standard but it is not perfect. The synthetic gold standard option allows for generating synthetic "lab-grown" graphs for precisely evaluating knowledge graph embeddings.

There are three flavors: Random graph generation (very expensive), constructed graph generation (in easy and hard), and ontology-based graph generation (recommended).

You can find the current version of the constructed gold standards the results directory.

Structure of the Results Folder and Terminology

Test Case Collection
The first directory order in the results directory is called test case collection in the implementation. Examples for a test case collection are tc01, tc02. A test case collection is a collection of multiple test case groups (see below). All test cases in a test case collection test for the same DL construct.

Test Case Group
A test case group is a domain-bound group of test cases. Examples for a test case group are people or movies.

Note that the synthetic results only have one test case group (e.g. synthetic_ontology). In the synthetic case, there is additionally a file graph.nt (the generated graph) and a directory dgl-ke-graph (the same graph in the DGL-KE format).

Size Group
A size is merely the size of the test case; more specifically, the number of positives and negatives, respectively.

Example: Structure of the DBpedia Gold Standard Directory

v1/
|
-- dbpedia/
   | 
   -- tc01/
      |
      -- books/
      |   |
      |   -- 50/ 
      |   |   |
      |   |   -- negatives.txt
      |   |   |
      |   |   -- positives.txt
      |   |   |
      |   |   -- train_test/
      |   |      |
      |   |      -- test.txt
      |   |      |
      |   |     -- train.txt
      |   |
      |    -- 500/
      |       |
      |       -- ...
      |
      -- cities/
         |
         -- ...

Example: Structure of the Synthetic Gold Standard Directory

v1/
|
-- synthetic_ontology/
   | 
   -- tc01/
   |   |
   |   -- synthetic_ontology/
   |      |
   |      -- 1000/ 
   |      |   |
   |      |   -- negatives.txt
   |      |   |
   |      |   -- positives.txt
   |      |   |
   |      |   -- train_test/
   |      |      |
   |      |      -- test.txt
   |      |      |
   |      |     -- train.txt
   |      |
   |      |
   |      -- graph.nt
   |      |
   |      -- ontology.nt
   |      |
   |      -- dgl-ke-graph/
   |
   -- tc02/
      |
      -- ...

Generator Command Line Interface (CLI)

The CLI can be used on to generate a gold standard without editing code in an IDE. You can call the help menu via -h/--help.

 -a,--analyze
 -b,--branchingFactor <arg>   Only valid for synthetic generators. The
                              parameter specifies the branching factor for
                              the ontology class tree.
 -c,--classes <arg>           Only valid for synthetic generators. The
                              parameter specifies the number of classes to
                              be used.
 -d,--directory <arg>         The test directory that shall be written or
                              analyzed (-a). If the directory shall be
                              written, it must not exist yet.
 -e,--edges <arg>             Only valid for synthetic generators. The
                              parameter specifies the number of edges to
                              be used.
 -h,--help                    Print this help message.
 -m,--maxTriples <arg>        Only valid for synthetic generators. The
                              parameter specifies the maximum number of
                              triples per node.
 -n,--nodeFactor <arg>        Only valid for synthetic generators. The
                              total nodes factor determines the maximum
                              number of nodes in a graph (totalNodesFactor
                              * nodesOfInterest).
 -q,--queries <arg>           The directory where the queries reside. If
                              parameter -q is not specified, the synthetic
                              query module is used.
 -s,--sizes <arg>             The sizes for the test cases, space
                              separated.
 -t,--timeout <arg>           Time out in seconds for queries.
 -tc,--tc_group <arg>         The test case group such as 'cities'; space
                              separated.
 -tcc,--tc_collection <arg>   The test case collection such as 'tc01';
                              space separated.

Development Remarks

  • java 17
  • maven project