The multi-modal version of DBP15K dataset comes from the EVA repository, and the folder pkls of DBP15K image features should be downloaded according to the guidance of EVA repository, and the downloaded folder pkls is placed in the data directory of this repository.

The word embedding we used is glove-6B, you can download it from glove, and unzip it into the data/embedding directory.

The original cross-KG datasets (FB15K-DB15K/YAGO15K) comes from MMKB, in which the image embeddings are extracted from the pre-trained VGG16. We use the image embeddings provided by MMKB and transform the data into the format consistent with DBP15K. The converted dataset can be downloaded from BaiduDisk (the password is stdt), and placed them in the data directory.

Here is the example of training MEAFE on DBP15K.

bash run_dbp15k.sh 0 42 zh_en
bash run_dbp15k.sh 0 42 ja_en
bash run_dbp15k.sh 0 42 fr_en

Here is the example of training MEAFE on FB15K_DB15K with different ratio seeds. Similarly, you can replace the parameter FB15K_DB15K with FB15K_YAGO15K to train FB15K-YAGO15K dataset.

bash run_mmkb_all.sh 0 42 FB15K_DB15K 0.2
bash run_mmkb_all.sh 0 42 FB15K_DB15K 0.5
bash run_mmkb_all.sh 0 42 FB15K_DB15K 0.8

Our codes are modified based on MCLEA, and we would like to thank their open-sourced work.