UQ-Neusoft-Health-Data-Science / ActiveEA

ActiveEA: Active Learning for Neural Entity Alignment

ActiveEA

Source code of paper "ActiveEA: Active Learning for Neural Entity Alignment", which has been published in EMNLP 2021.

Steps of reproducing the experiments:

• Step 1: Download and unzip the repo.
• Step 2: If you need to run our strategies on RDGCN, you need to download the open word embedding file from wiki-news-300d-1M.vec and put the unzipped file under dataset/. Otherwise, skip this step (the size of unzipped word embedding file will be 2.26GB).
• Step 3: Install conda environment. cd to your project directory firstly. Then, create the environment using command below.

conda env create -f environment.yml

Then, activate the environment via conda activate al4ea, and install more packages using the following commands

conda install -c conda-forge graph-tool==2.29
pip install igraph
pip install python-Levenshtein
pip install gensim==4.0.1

• Step 4: Configure settings. The scripts to run are under scripts/run_strategies/ The default settings are set in task_settings.sh. Before you run any script, set proj_dir in the setting file firstly.

• Step 5: Run scripts:

  • For trials: customizing script task_runner_trial.sh.
  • Run experiments about the "overall performance on 15K data": task_runner_overall_perf.sh.
  • Run experiments about the "overall performance on 15K data": task_runner_overall_perf_100k.sh.
  • Run experiments about the "effect of bachelors": task_runner_effect_of_bachelor_percent.sh.
  • Run experiments about the "effectiveness of bachelor recognizer": intermediate results have been saved with the generated dataset of AL process.
  • Run experiments about the "sensitivity of parameters": task_runner_effect_of_alpha.sh and task_runner_effect_of_batchsize.sh.

The generated datasets by different AL strategies will be saved to dataset/ with naming pattern like dataset/${seed}/${task_group}/${dataset_name}/${strategy_name}. The evaluation results on test set will be saved to output/results/.

Citation

If you re-use our code for your paper, please kindly cite our paper:

@inproceedings{DBLP:conf/emnlp/LiuSZHZ21,
  author    = {Bing Liu and
               Harrisen Scells and
               Guido Zuccon and
               Wen Hua and
               Genghong Zhao},
  editor    = {Marie{-}Francine Moens and
               Xuanjing Huang and
               Lucia Specia and
               Scott Wen{-}tau Yih},
  title     = {ActiveEA: Active Learning for Neural Entity Alignment},
  booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural
               Language Processing, {EMNLP} 2021, Virtual Event / Punta Cana, Dominican
               Republic, 7-11 November, 2021},
  pages     = {3364--3374},
  publisher = {Association for Computational Linguistics},
  year      = {2021},
  url       = {https://doi.org/10.18653/v1/2021.emnlp-main.270},
  doi       = {10.18653/v1/2021.emnlp-main.270},
  timestamp = {Thu, 20 Jan 2022 10:02:11 +0100},
  biburl    = {https://dblp.org/rec/conf/emnlp/LiuSZHZ21.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Acknowledgement

We implement the neural EA models by customizing source code of OpenEA.