data_process

KG Construction

Original Data Introduction and Preparation

For each dataset, we prepare the following original data:

1. class file class.json: the seen and unseen class information in each dataset, including WordNet ID and literal name;

2. attribute file

   • attribute.txt: attribute information, including custom ID and literal name;
   • attribute_hierarchy.owl & attribute_group.json: the categorization information of class attributes;

3. attribute annotation file

   • AwA: binaryAtt_splits.mat and att_splits.mat, download from here and put it into the folder ori_data/AwA;
   • ImNet-A/O: class_attribute.json

4. ConceptNet (5.7): the full set can be download from here, its English subset can be extracted by running conceptnet_en_subgraph.py

Data Processing

For each dataset, we run the following scripts to construct KG:

1. class_hierarchy.py: build hierarchical structure of classes as the backbone of KG, output class_hierarchy_triples.txt

2. attribute_hierarchy.py: build hierarchical structure of attributes, output attribute_hierarchy_triples.txt

3. class_attribute_awa.py & class_attribute_imagenet.py: build attribute triples, output class_attribute_triples.txt

4. literal.py: add literal information of nodes in graph, output literals.txt

5. Link to ConceptNet

   • conceptnet_alignment_extraction.py: align classes and attributes to conceptnet entities and extract their one-hop neighbor subgraph;
   • conceptnet_entity_alignment.py: save the aligned pairs, output sameAs_triples.txt;
   • conceptnet_repeat_check.py: remove repetitive triples, output conceptnet_triples_filter.txt

6. Set disjointness semantics

   • disjointness_classes.py: disjointness between different classes, output disjoint_cls_cls_triples.txt
   • disjointness_cls_atts.py: disjointness between classes and attributes, output disjoint_cls_att_triples.txt

7. Run output2CSV.py to save KGs to CSV file. Note we set different parameters to output KGs with different semantic settings. Taking AwA as an example:

   • generate KG with all semantics by running python output2CSV.py --dataset AwA --all
   • generate KG with semantics of class hierarchy by running python output2CSV.py --dataset AwA --cls_hie
   • generate KG with semantics of class hierarchy and attribute hierarchy by running python output2CSV.py --dataset AwA --cls_hie --att_hie

8. Run output2RDF.py to save KGs with Turtle and XML files.