MKGFormer
Code for the SIGIR 2022 paper "Hybrid Transformer with Multi-level Fusion for Multimodal Knowledge Graph Completion"
- ❗NOTE: We provide some KGE baselines at OpenBG-IMG.
- ❗NOTE: We release a new MKG task "Multimodal Analogical Reasoning over Knowledge Graphs (ICLR'2023)" at MKG_Analogy.
Model Architecture
Illustration of MKGformer for (a) Unified Multimodal KGC Framework and (b) Detailed M-Encoder.
Requirements
To run the codes (Python 3.8), you need to install the requirements:
pip install -r requirements.txt
Data Preprocess
To extract visual object images int MNER and MRE tasks, we first use the NLTK parser to extract noun phrases from the text and apply the visual grouding toolkit to detect objects. Detailed steps are as follows:
- Using the NLTK parser (or Spacy, textblob) to extract noun phrases from the text.
- Applying the visual grouding toolkit to detect objects. Taking the twitter2017 dataset as an example, the extracted objects are stored in
twitter2017_aux_images
. The images of the object obey the following naming format:id_pred_yolo_crop_num.png
, whereid
is the order of the raw image corresponding to the object,num
is the number of the object predicted by the toolkit. (id
is doesn't matter.) - Establishing the correspondence between the raw images and the objects. We construct a dictionary to record the correspondence between the raw images and the objects. Taking
twitter2017/twitter2017_train_dict.pth
as an example, the format of the dictionary can be seen as follows:{imgname:['id_pred_yolo_crop_num0.png', 'id_pred_yolo_crop_num1.png', ...] }
, where key is the name of raw images, value is a List of the objects (Note that intrain/val/test.txt
, text and raw image have a one-to-one relationship, so theimgnae
can be used as a unique identifier for the raw images).
The detected objects and the dictionary of the correspondence between the raw images and the objects are available in our data links.
Data Download
The datasets that we used in our experiments are as follows:
-
Twitter2017
You can download the twitter2017 dataset from Google Drive.
For more information regarding the dataset, please refer to the UMT repository.
-
MRE
The MRE dataset comes from MEGA, many thanks.
You can download the MRE dataset with detected visual objects from Google Drive or using following command:
cd MRE wget 120.27.214.45/Data/re/multimodal/data.tar.gz tar -xzvf data.tar.gz
-
MKG
-
FB15K-237-IMG
You can download the image data of FB15k-237 from mmkb which provides a list of image URLs, and refer to more information of description of entity from kg-bert repositories.
- ❗NOTE: we have found a severe bug in the code of data preprocessing for FB15k-237-IMG, which leads to the unfair performance comparison; we have updated the performance in arxiv and released the checkpoints (The model trained with/without the severe bug).
-
WN18-IMG
Entity images in WN18 can be obtained from ImageNet, the specific steps can refer to RSME. the RSME repository.
-
We also provide additional network disk links for multimodal KG data (Images) at GoogleDrive or Baidu Pan with extraction (code:ilbd).
The expected structure of files is:
MKGFormer
|-- MKG # Multimodal Knowledge Graph
| |-- dataset # task data
| |-- data # data process file
| |-- lit_models # lightning model
| |-- models # mkg model
| |-- scripts # running script
| |-- main.py
|-- MNER # Multimodal Named Entity Recognition
| |-- data # task data
| | |-- twitter2017
| | | |-- twitter17_detect # rcnn detected objects
| | | |-- twitter2017_aux_images # visual grounding objects
| | | |-- twitter2017_images # raw images
| | | |-- train.txt # text data
| | | |-- ...
| | | |-- twitter2017_train_dict.pth # {imgname: [object-image]}
| | | |-- ...
| |-- models # mner model
| |-- modules # running script
| |-- processor # data process file
| |-- utils
| |-- run_mner.sh
| |-- run.py
|-- MRE # Multimodal Relation Extraction
| |-- data # task data
| | |-- img_detect # rcnn detected objects
| | |-- img_org # raw images
| | |-- img_vg # visual grounding objects
| | |-- txt # text data
| | | |-- ours_train.txt
| | | |-- ours_val.txt
| | | |-- ours_test.txt
| | | |-- mre_train_dict.pth # {imgid: [object-image]}
| | | |-- ...
| | |-- vg_data # [(id, imgname, noun_phrase)], not useful
| | |-- ours_rel2id.json # relation data
| |-- models # mre model
| |-- modules # running script
| |-- processor # data process file
| |-- run_mre.sh
| |-- run.py
How to run
-
MKG Task
- First run Image-text Incorporated Entity Modeling to train entity embedding.
cd MKG bash scripts/pretrain_fb15k-237-image.sh
- Then do Missing Entity Prediction.
bash scripts/fb15k-237-image.sh
-
MNER Task
To run mner task, run this script.
cd MNER bash run_mner.sh
-
MRE Task
To run mre task, run this script.
cd MRE bash run_mre.sh
Acknowledgement
The acquisition of image data for the multimodal link prediction task refer to the code from https://github.com/wangmengsd/RSME, many thanks.
Papers for the Project & How to Cite
If you use or extend our work, please cite the paper as follows:
@inproceedings{DBLP:conf/sigir/ChenZLDTXHSC22,
author = {Xiang Chen and
Ningyu Zhang and
Lei Li and
Shumin Deng and
Chuanqi Tan and
Changliang Xu and
Fei Huang and
Luo Si and
Huajun Chen},
editor = {Enrique Amig{\'{o}} and
Pablo Castells and
Julio Gonzalo and
Ben Carterette and
J. Shane Culpepper and
Gabriella Kazai},
title = {Hybrid Transformer with Multi-level Fusion for Multimodal Knowledge
Graph Completion},
booktitle = {{SIGIR} '22: The 45th International {ACM} {SIGIR} Conference on Research
and Development in Information Retrieval, Madrid, Spain, July 11 -
15, 2022},
pages = {904--915},
publisher = {{ACM}},
year = {2022},
url = {https://doi.org/10.1145/3477495.3531992},
doi = {10.1145/3477495.3531992},
timestamp = {Mon, 11 Jul 2022 12:19:20 +0200},
biburl = {https://dblp.org/rec/conf/sigir/ChenZLDTXHSC22.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}