Semantics Driven Embedding Learning for Effective Entity Alignment

Dependencies

• python: 3.7
• pytorch: 1.3.1
• transformers: 4.2.2
• tqdm: 4.56.0

Installation

We recommend creating a new conda environment to install the dependencies and run SDEA:

conda create -n SDEA python=3.7
conda activate SDEA
conda install pytorch-gpu=1.3.1
pip install transformers==4.2.2

Preparation

The structure of the project is listed as follows:

SDEA/
├── src/: The soruce code of SDEA. 
├── data/: The datasets. 
│   ├── DBP15k/: The downloaded DBP15K benchmark. 
│   │   ├── fr_en/
│   │   ├── ja_en/
│   │   ├── zh_en/
│   ├── entity-alignment-full-data/: The downloaded SRPRS benchmark. 
│   │   ├── en_de_15k_V1/
│   │   ├── en_fr_15k_V1/
│   │   ├── dbp_wd_15k_V1/
│   │   ├── dbp_yg_15k_V1/
├── pre_trained_models/: The pre-trained transformer-based models. 
│   ├── bert-base-multilingual-uncased: The model used in our experiments.
│   │   ├── config.json
│   │   ├── pytorch_model.bin
│   │   ├── tokenizer.json
│   │   ├── tokenizer_config.json
│   │   ├── vocab.txt
│   ├── ......

Datasets

• SRPRS: https://github.com/nju-websoft/RSN/raw/master/entity-alignment-full-data.7z

• DBP15K: http://ws.nju.edu.cn/jape/data/DBP15k.tar.gz Service unavailable now. Please download from Google Drive.

1. Download the datasets and unzip them into "SDEA/data".

2. Preprocess the datasets.

cd src
python DBPDSPreprocess.py
python SRPRSPreprocess.py

Pre-trained Models

The pre-trained models of transformers library can be downloaded from https://huggingface.co/models. We use bert-base-multilingual-uncased in our experiments.

Please put the downloaded pre-trained models into "SDEA/pre_trained_models".

How to Run

bash run_dbp15k.sh
bash run_SRPRS.sh

Citation

If you find our work useful, please kindly cite the following paper:

@inproceedings{SDEA,
  author    = {Ziyue Zhong and
               Meihui Zhang and
               Ju Fan and
               Chenxiao Dou},
  title     = {Semantics Driven Embedding Learning for Effective Entity Alignment},
  booktitle = {{ICDE}},
  pages     = {2127--2140},
  publisher = {{IEEE}},
  year      = {2022},
  doi       = {10.1109/ICDE53745.2022.00205}
}