Code and datasets for the ICLR2023 paper "Multimodal Analogical Reasoning over Knowledge Graphs "
The structure of data files is as follows:
MKG_Analogy
|-- MarT
| |-- dataset # data
| | |-- MarKG # knowledge graph data
| | | |-- entity2text.txt # entity_id to entity_description
| | | |-- entity2textlong.txt # entity_id to longer entity_description
| | | |-- relation2text.txt # relation_id to relation_description
| | | |-- relation2textlong.txt # relation_id to longer relation_description
| | | |-- wiki_tuple_ids.txt # knowledge triplets with (head_id, rel_id, tail_id) format
| | |-- MARS # analogical reasoning data
| | | |-- images # the image data
| | | |-- analogy_entities.txt # the analogical entities
| | | |-- analogy_entity_to_wiki_qid.txt # the ids of analogical entities
| | | |-- analogy_relations.txt # the analogical relations
| | | |-- dev.json # analogical reasoning data for validation
| | | |-- test.json # analogical reasoning data for testing
| | | |-- train.json # analogical reasoning data for training
The image data can be downloaded through this link with extraction code (7hoc) and placed on MarT/dataset/MARS/images
.
The format of the knowledge triplet data (wiki_tuple_ids.txt) is as below:
Q15026 P276 Q107
Q34266 P47 Q30
Q317557 P140 Q5043
Q8686 P910 Q7214221
In each line, the entity ids start with Q
and relation ids start with P
.
The analogical reasoning data is stored in json format. For each analogical reasoning instance, we provide an analogical example and an analogical question entity to acquire the analgical answer entity, the data format is as below:
{
"example": ["Q14536140", "Q581459"],
"question": "Q50000",
"answer": "Q202875",
"relation": "P828",
"mode": 0
}