search

OreOZhao / MoSE4MKGC

Code for "MoSE: Modality Split and Ensemble for Multimodal Knowledge Graph Completion", EMNLP, 2022.
27 stars 2 forks source link

readme

MoSE4MKGC

Code for paper "MoSE: Modality Split and Ensemble for Multimodal Knowledge Graph Completion", EMNLP, 2022.

Dataset Download

You can download dataset FB15K-237, WN18 from kbc, and WN9 from IKRL.

The descriptions of entities can be downloaded from KG-BERT.

The images of entities can be downloaded from MMKB, RSME, or MKGformer.

Dataset Preprocess

For triples data, the data could be preprocessed by src/process_datasets.py.

For text data, we used bert-base-uncased model from transformers as encoder. We use the mean aggregation of token hidden states as text feature of entity. The text encoder is in src/text_encoder.py.

For image data, we used pytorch_pretrained_vit model as encoder, which is the same as RSME. The image encoder is in src/img_encoder.py.

For the missing descriptions or images, we impute the feature with np.random.normal vector.

After extracting features from fixed encoders, we save the text and image features of entities in a pickle file and save the file in data/DATASET_NAME/.

The expected project structure is:

MoSE4MKGC
 |-- src
 |-- src_data           # downloaded triples of dataset
 |    |-- DATASET_NAME
 |    |    |-- train
 |    |    |-- valid
 |    |    |-- test
 |    |    |-- ...      # other files
 |-- data
 |    |-- DATASET_NAME
 |    |    |-- ...      # files from process_datasets.py
 |    |    |-- YOUR_TEXT_FEATURE.pickle
 |    |    |-- YOUR_IMG_FEATURE.pickle

How to run

After preprocessing, you could run src/learner.py to train the modality split KGs. 

cd src
CUDA_VISIBLE_DEVICES='1' python learn.py --model ComplExMDR --ckpt_dir YOUR_CKPT_DIR --dataset WN9 --early_stopping 10 --fusion_dscp True --fusion_img True --modality_split True --img_info PATH_TO_YOUR_IMG_FEATURE.pickle --dscp_info PATH_TO_YOUR_TEXT_FEATURE.pickle

Note that the displayed metrics during training are more like an upper bound of modality split KGs performance, which is not the final prediction performance.

Then you can run src/boosting_inference.py or src/meta_learner.py to ensemble the modality split predictions to get final predictions and performance.

MoSE-AI

cd src
python boosting_inference.py --model_path YOUR_MODEL_PATH --dataset DATASET_NAME --boosting False

MoSE-BI

cd src
python boosting_inference.py --model_path YOUR_MODEL_PATH --dataset DATASET_NAME --boosting True

MoSE-MI

cd src
python meta_learner.py --model_path YOUR_MODEL_PATH --dataset DATASET_NAME

Acknowledgement

Thanks to the support of RSME for the image acquisition and code open source.

Citation

Please cite the following paper as reference if you find our work useful.


@inproceedings{zhao-etal-2022-mose,
    title = "{M}o{SE}: Modality Split and Ensemble for Multimodal Knowledge Graph Completion",
    author = "Zhao, Yu  and
      Cai, Xiangrui  and
      Wu, Yike  and
      Zhang, Haiwei  and
      Zhang, Ying  and
      Zhao, Guoqing  and
      Jiang, Ning",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.719",
    pages = "10527--10536"
}