wbsg-uni-mannheim / wdcproducts

This repository contains the code and data download links to reproduce building the WDC Products Benchmark.

WDC Products: A Multi-Dimensional Entity Matching Benchmark

This repository contains the code and data download links to reproduce building the WDC Products benchmark. The benchmark files are available for direct download here. A paper describing the benchmark will be presented at the EDBT2024 conference. A preprint of the paper is available here.

• Requirements

  Anaconda3

  Please keep in mind that the code is not optimized for portable or even non-workstation devices. Some of the scripts may require large amounts of RAM (64GB+) and GPUs. It is advised to use a powerful workstation or server when experimenting with some of the larger files.

  The code has only been used and tested on Linux (CentOS) servers.

• Building the conda environment

  To build the exact conda environment used for the creation of WDC Products, navigate to the project root folder where the file environment.yml is located and run conda env create -f environment.yml

• Downloading the raw data files

  Navigate to the src/data/ folder and run python download_datasets.py to automatically download the source files into the correct locations. You can find the data at data/interim/wdc-lspc/corpus/

• Building WDC Products

  To reproduce building WDC Products, run the following notebooks in order. Please keep in mind that some notebooks may run for multiple hours.

  1. notebooks/processing/benchmark2020/langdetect-and-clean.ipynb
  2. notebooks/processing/benchmark2020/dbscan-clustering.ipynb
  3. notebooks/processing/benchmark2020/generate-sets-final.ipynb

• Experiments

  You need to install the project as a package. To do this, activate the environment, navigate to the root folder of the project, and run pip install -e .

  • Preparing the data:

  To prepare the data for the experiments, run the first script below and then any of the others to prepare the data for the respective experiments. Make sure to navigate to the respective folders first. If you did not run the previous benchmark creation steps, you can obtain the needed data files on the WDC Products page:

  • src/processing/preprocess/preprocess_wdcproducts.py

  • src/processing/contrastive/prepare-data.py

  • src/processing/process-magellan/process_to_magellan.py

  • src/processing/process-wordcooc/process-to-wordcooc.py

  • src/processing/process-wordocc/process-to-wordocc-multi.py

  To prepare data for Ditto and HierGAT input run

  • preprocess_data_for_ditto_hiergat.py

  • Running the experiments:

    • Ditto: First install the Ditto environment from the ditto/ditto_env.yml then run ditto/all_runs.py

    • HierGAT: First install the HierGAT environment from the hiergat/hiergat_env.yml then run hiergat/all_runs.py

    • Magellan: Navigate to src/models/magellan/ and run the script run_magellan.py

    • Word Coocurrence: Navigate to src/models/wordcooc/ and run the script run_wordcooc.py

    • Word Occurrence: Navigate to src/models/wordocc and run the script run_wordocc_multi.py

    • Transformer:

      Navigate to src/contrastive/

      To fine-tune a Transformer, run any of the fine-tuning scripts, e.g. for pair-wise:

      CUDA_VISIBLE_DEVICES="GPU_ID" bash lspc/run_finetune_baseline.sh roberta-base True 64 5e-05 wdcproducts80cc20rnd000un large

      You need to specify model, usage of gradient checkpointing, batch size, learning rate, dataset and development size as arguments here.

      Analogously for fine-tuning a multi-class Transformer:

      CUDA_VISIBLE_DEVICES="GPU_ID" bash lspc/run_finetune_baseline_multi.sh roberta-base True 64 5e-05 wdcproductsmulti80cc20rnd000un large

    • R-SupCon:

      Navigate to src/contrastive/

      • Contrastive Pre-training:

        To run contrastive pre-training use e.g.

        CUDA_VISIBLE_DEVICES="GPU_ID" bash lspc/run_pretraining.sh roberta-base True 1024 5e-05 0.07 wdcproducts80cc20rnd000un large

        You need to specify model, usage of gradient checkpointing, batch size, learning rate, temperature, dataset and development size as arguments here.

      • Cross-entropy Fine-tuning:

        Finally, to use the pre-trained models for fine-tuning, run any of the fine-tuning scripts, e.g. for pair-wise:

        CUDA_VISIBLE_DEVICES="GPU_ID" bash lspc/run_finetune_siamese.sh roberta-base True 1024 5e-05 0.07 frozen wdcproducts80cc20rnd000un large

        Analogously for fine-tuning multi-class R-SupCon:

        CUDA_VISIBLE_DEVICES="GPU_ID" bash lspc/run_finetune_multi.sh roberta-base True 1024 5e-05 0.07 frozen wdcproductsmulti80cc20rnd000un large

  Result files can subsequently be found in the reports folder.

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience