KGMEL: Knowledge Graph-Enhanced Multimodal Entity Linking

Official Repository for our paper "KGMEL: Knowledge Graph-Enhanced Multimodal Entity Linking" (SIGIR 2025 Short)

Setup

0. Data Preparation

Our dataset is based on the MIMIC repository.
You need to download the following files from Our GoogleDrive:

• triple.tar.gz - Knowledge graph triples
• mention_image.tar.gz - Mention images
• kb_image.tar.gz - Knowledge base entity images

Then, extract the files:

# Extract mention images for datasets
tar -xf mention_image.tar.gz -C data/dataset/

# Extract Knowledge Base images
tar -xf kb_image.tar.gz -C data/KB/

# Extract Knowledge Graph triples
tar -xf triple.tar.gz -C data/KB/

1. Install Dependencies

Install the required Python packages:

pip install -r requirements.txt

2. Set Keys and Tokens

Set up the necessary API keys and tokens:

# Set OpenAI API key
export OPENAI_API_KEY="your_openai_api_key"
# Set Hugging Face token
export HUGGINGFACE_TOKEN="your_huggingface_token"
# Set Hugging Face cache directory (optional)
export HUGGINGFACE_HUB_CACHE="/path/to/huggingface/cache"
# Set the WANDB_API_KEY environment variable
export WANDB_API_KEY="your_wandb_api_key"
# Login to Hugging Face (alternative method)
# huggingface-cli login
# Login to Weights & Biases (alternative method)
# wandb login

You can add these to your .bashrc or .bash_profile for persistence, or include them in your run script for convenience.

3. Run the Code

Execute the run script:

./run.sh

Structure

KGMEL/
├── checkpoints/         # Directory for saving trained models
├── data/
│   ├── dataset/        # Contains the 3 MEL datasets (WikiMEL, WikiDiverse, RichpediaMEL)
│   │   ├── image/      # Contains extracted image datasets from mention_image.tar
│   │   ├── mapping/    # Contains mapping files for dataset processing
│   │   ├── RichpediaMEL.json
│   │   ├── WikiDiverse.json
│   │   └── WikiMEL.json
│   └── KB/              # Contains Knowledge Base data with including KG triples
│       ├── image/       # Contains extracted image datasets from kb_image.tar
│       ├── PID2Label.tsv
│       ├── QID2Label.tsv
│       ├── Triples-RichpediaMEL.tsv  # These triple files come from triples.tar.gz
│       ├── Triples-WikiDiverse.tsv
│       └── Triples-WikiMEL.tsv
├── embedding/           # Directory for pre-computed embeddings
├── module/
│   ├── generate.py      # Triple Generation module
│   ├── retrieve.py      # Candidate Entity Retrieval module
│   └── rerank.py        # Entity Reranking module
├── output/              # Output directory for results
├── utils/
│   ├── dataloader.py
│   ├── embedding_processor.py
│   ├── encoder.py
│   ├── evaluate.py
│   ├── train.py
│   ├── triple_filtering.py
│   └── triple_parser.py
├── main.py              # Main code 
├── requirements.txt     # Dependencies
└── run.sh               # Execution script

Citation

@inproceedings{kim2025kgmel,
  title={KGMEL: Knowledge Graph-Enhanced Multimodal Entity Linking},
  author={Kim, Juyeon and Lee, Geon and Kim, Taeuk and Shin, Kijung},
  booktitle={Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2025}
}