🫙 JAR: Join-Aware Multi-Table Retrieval

If you find our code, or the paper helpful, please cite the paper

@article{chen2024table,
  title={Is Table Retrieval a Solved Problem? Join-Aware Multi-Table Retrieval},
  author={Chen, Peter Baile and Zhang, Yi and Roth, Dan},
  journal={arXiv preprint arXiv:2404.09889},
  year={2024}
}

Overview

• Query decomposition (decomp.py)
• Table-table compatibility scores (compatibility.py)
• Mixed-integer linear program (MIP) to retrieve the optimal set of tables (ilp.py)
• Baselines
  • DPR: Contriever (contriever.py)
  • DTR: TAPAS (tapas.py)

Setup

To execute the MIP program, we recommend using the Gurobi solver for significantly faster execution.

Table-table compatibility scores generated from compatibility.py and query decomposition generated from decomp.py for Spider and Bird have been pre-computed and saved under data/.

The dev_database for Bird and Spider as well as the fine-tuned TAPAS-large checkpoints can be found in this Google drive folder.

You need to first execute contriever.py or tapas.py to obtain the coarse-grained relevance score and execute contriever.py to obtain fine-grained relevance score before you can run ilp.py.

Files with

parser = argparse.ArgumentParser()
parser.add_argument('-p', '--partition', type=int)
args = parser.parse_args()

support parallel execution. num_partitions specifies the the number of jobs executed in parallel. To run these files (e.g., ilp.py), you can use, for example,

python ilp.py -p 0 & python ilp.py -p 1 & ...

Contact

If you have any questions or feedback, please send an email to [email protected].