Rethinking Table Retrieval from Data Lakes

Table retrieval from data lakes has recently become important for many downstream tasks, including data discovery and table question answering. Existing table retrieval approaches estimate each table's relevance to a particular information need and return a ranking of the most relevant tables. This approach is not ideal since (1) the returned tables often include irrelevant data and (2) the required information may be scattered across multiple tables. To address these issues, we propose the idea of fine-grained structured table retrieval and present our vision of R2D2, a system which slices tables into small tiles that are later composed into a structured result that is tailored to the user-provided information need. An initial evaluation of our approach demonstrates how our idea can improve table retrieval and relevant downstream tasks such as table question answering.

Please check out our paper and cite our work:

@inproceedings{DBLP:conf/aidm/BodensohnB24,
    author       = {Jan{-}Micha Bodensohn and
                  Carsten Binnig},
    title        = {Rethinking Table Retrieval from Data Lakes},
    booktitle    = {Proceedings of the Seventh International Workshop on Exploiting Artificial
                  Intelligence Techniques for Data Management, aiDM 2024, Santiago,
                  Chile, 14 June 2024},
    pages        = {2:1--2:5},
    publisher    = {{ACM}},
    year         = {2024},
    url          = {https://doi.org/10.1145/3663742.3663972},
    doi          = {10.1145/3663742.3663972}
}

Setup

Make sure you have Conda installed.

Create a new Conda environment, activate it, and add the project to the Python path:

eval "$(conda shell.bash hook)"
conda env create -f environment-cuda.yml  # use environment.yml if you don't have CUDA
conda activate r2d2
export PYTHONPATH=${PYTHONPATH}:./

Datasets

Open-WikiTable

To prepare the Open-WikiTable dataset, run:

python scripts/lakes/download_openwikitable.py
python scripts/lakes/create_openwikitable_lakes.py

Spider

To prepare the Spider dataset, you first have to manually obtain the dataset.

Download the ZIP archive from here and unpack its contents into data/spider/download/.

Then run:

python scripts/lakes/create_spider_lakes.py  # runs for several hours

Experiments

Dataset Exploration

The exploration scripts generate histograms about the dataset statistics (e.g., tables per lake, cells per table...) and can also plot the ground truths.

To explore the datasets, run:

bash scripts/explore/explore.sh

The results will be placed in data/<dataset>/explore.

Slicing

To evaluate the slicing, run:

bash scripts/slicing/slicing.sh  # runs multiple hours

The results will be placed in data/<dataset>/slicing.

Table Question Answering

To evaluate the table question answering, run:

bash scripts/tqa/tqa.sh  # runs multiple hours

The results will be placed in data/<dataset>/tqa.

To replicate our results you must obtain the data/openai_cache directory from us.