Skip to content

Latest commit

 

History

History
166 lines (110 loc) · 3.99 KB

README_EN.md

File metadata and controls

166 lines (110 loc) · 3.99 KB

OpenBG-IMG Benchmark

OpenBG-IMG is a multi-modal dataset in the field of e-commerce. This benchmark implements a variety of knowledge graph embedding models for link prediction, which is used to generate evaluation results of CCKS2022 Knowledge Processing and Application Evaluation for Digital Commerce Task 3: Multimodal Commodity Knowledge Graph Link Prediction. Please submit the results to TIANCHI platform.

Requirements

Use the following code to configure the environment.

pip install -r requirements.txt

Dataset

Please download the dataset from TIANCHI to the directory ./data/.

data
 |-- OpenBG-IMG
 |    |-- images            # Set of images
 |    |    |-- ent_xxxxxx   # Images of the entity
 |    |    |-- ...
 |    |-- train.tsv         # Training set
 |    |-- test.tsv          # Test set

The statistics of OpenBG-IMG：

Dataset # Ent # Rel # Train # Dev # Test
OpenBG-IMG 27,910† 136 230,087 5,000 14,675

†: there are 14,718 multi-modal entities in OpenBG-IMG.

Check the data

$ head -n 3 train.tsv
ent_021198	rel_0031	ent_017656
ent_008185	rel_0092	ent_025949
ent_005940	rel_0080	ent_020805

Quick Start

TransE & TransH & TransE & DistMult & ComplEx

These models refer to the implement of OpenKE.

  • Compile C++ files
    cd [The root directory of model]
    bash scripts/make.sh
  • Preprocess data
    bash scripts/prepro.sh
  • Train model and predict results saved to ./results/result.tsv
    bash scripts/train.sh

TuckER

The model refers to the implement of TuckER.

  • Preprocess data
    bash scripts/prepro.sh
  • Train model and predict results saved to ./results/result.tsv
    bash scripts/train.sh

TransAE

The model refers to the implement of the TransE model in OpenKE and the process of images in TransAE.

  • Compile C++ files
    cd TransAE
    bash scripts/make.sh
  • Preprocess data
    bash scripts/prepro.sh
  • Get representations and encodings of images
    bash scripts/visual_emb.sh
  • Train model and predict results saved to ./results/result.tsv
    bash scripts/train.sh

RSME

The model refers to the implement of RSME.

  • Get representations and encodings of images
    cd RSME
    bash scripts/visual_emb.sh
  • Preprocess data
    bash scripts/prepro.sh
  • Train model and predict results saved to ./results/result.tsv
    bash scripts/train.sh

Experiments

Model HIT@1 HIT@3 HIT@10 MR MRR
TransE 0.150 0.387 0.647 118 0.315
TransH 0.129 0.525 0.743 112 0.357
TransD 0.137 0.532 0.746 110 0.364
DistMult 0.060 0.157 0.279 524 0.139
ComplEx 0.143 0.244 0.371 782 0.221
TuckER 0.497 0.690 0.820 1473 0.611
TransAE 0.274 0.489 0.715 36.1 0.421
RSME 0.485 0.687 0.838 72.1 0.607

Acknowledgements

Thanks for the following works!

More related works

Multi-modal Knowledge Graph completion: MKGFormer(https://github.com/zjunlp/MKGformer/)