DMRAL

This repository contains the code and benchmarks for the paper Decomposition-Driven Multi-Table Retrieval and Reasoning for Numerical Question Answering.

---

📦 Benchmarks

We release two benchmarks to support evaluation:

• SpiderWild
• BirdWild

Each benchmark adapts an existing text-to-SQL dataset into a more realistic setting for multi-table question answering over large-scale table collections.

Setup

• Table collection:
  Download the benchmark files and place them in a tables subdirectory within the corresponding dataset folder.

• Questions and Labels:
  The associated questions, relevant tables, and answers are located in the dataset/label directory.

• Trained models: Download Coverage Scoring Function Models and put them under code/checkpoint

• LLM config: Modify "api_key" in model_config.py located in the code/utils directory.

---

🚀 Running the DMRAL Framework

1. Installation

Create and activate a conda environment, then install all required Python dependencies:

conda create -n dmral_env python=3.8 -y
conda activate dmral_env
pip install -r requirements.txt

2. Running the Full Pipeline

Note: Before running the bash.sh, make sure to update the dataset_name variable by speficifying the dataset name (e.g., "spiderwild").

cd code
bash bash.sh

Contact

If you have any issue, please contact feng.luo@student.rmit.edu.au.