QGpT: Improving Table Retrieval with Question Generation from Partial Tables

This repository contains the source code, corpora, and model prompts for the paper:
📄 Improving Table Retrieval with Question Generation from Partial Tables (TRL Workshop @ ACL 2025)

We propose QGpT, a simple yet effective framework that improves open-domain table retrieval by generating synthetic questions from partial tables.

---

🆕 Updates

2026/02/03

🎉 Latest Research at The Web Conference 2026 (WWW 2026)! Check out our two new papers on this topic:

• 📄 CGPT: Cluster-Guided Partial Tables with LLM-Generated Supervision for Table Retrieval | GitHub
• 📄 STAR: Semantic Table Representation with Header-Aware Clustering and Adaptive Weighted Fusion | GitHub

2025/10/30

We have released Milvus-Lite implementation for local vector database operations:

• ✅ Complete Python scripts for embedding and indexing table corpora
• ✅ Search and evaluation pipeline with Recall@k metrics
• ✅ Jupyter notebooks with interactive examples
• ✅ Conda environment configuration (Milvus.yml)

This allows you to reproduce our experiments locally without requiring a Milvus server deployment.

2025/08/08

We have released a original Question to gold-table datasets that we used to test recall@k:

➡️ Hugging Face Dataset: cc3374twa/QGPT

This dataset contains question-table pairs from the original datasets, and is structured into multiple subsets (e.g., E2E-WTQ).

To load the dataset using datasets:

from datasets import load_dataset

# Load the E2E-WTQ subset
dataset = load_dataset("cc3374twa/QGPT", name="E2E-WTQ", split="test")

🗂️ Repository Contents

• 📁 Corpora/
  Table corpora used in each experimental table in the paper. Each subfolder corresponds to one experiment section:

  • Table1_mimo_table_length_variation/
  • Table3_mimo_en_table_representation/
  • Table5_Single_Table_Retrieval/
  • Table6_Multi_Table_Retrieval/
  • Table7_OTTQA/

• 📁 prompt/
  Prompt templates for question generation and query decomposition (MMQA).

  • llama3-8b-Instruct_header_extract_and_QG.txt
  • llama3-8b-Instruct_QG_only.txt
  • MMQA_query_decomposition.txt

---

📦 Dataset Construction

The table corpora under Corpora/ are preprocessed and embedded based on our proposed method.
They are indexed using either:

• Milvus → https://milvus.io
• RAGatouille → https://github.com/AnswerDotAI/RAGatouille

🗄️ Milvus Implementation

We provide Python scripts for working with Milvus-Lite (single-file vector database):

• 📁 embedding_db/pymilvus/

  • create_pymilvus_db.py - Create Milvus collections from table corpora
  • pymilvus_embedding.py - Embed and insert data using BAAI/bge-m3 model

• 📁 evaluation/

  • search_pymilvus.py - Search queries in Milvus collections
  • evaluation.py - Evaluate retrieval performance with Recall@k metrics

• 📓 Jupyter Notebooks:

  • Milvus.ipynb - Complete workflow for Milvus database creation and embedding
  • E2EWTQ_Milvus_search.ipynb - E2E-WTQ dataset search example
  • export_milvus.py - Export metadata from Milvus collections

• ⚙️ Environment:

  • Milvus.yml - Conda environment specification with all dependencies

---

🧪 Reproducibility

• All table corpora are constructed based on the same datasets used in the paper.
• Each folder maps to the exact experimental tables (e.g., Table 1, Table 5).

⚠️ If any released data differs from what’s reported in the paper due to human error, please contact us at cc3374twa@gmail.com.

---

🚀 Quick Start

Setup Milvus Environment

# Create conda environment from yml file
conda env create -f Milvus.yml
conda activate Milvus

Create Vector Database

cd embedding_db/pymilvus
python create_pymilvus_db.py  # Create collections
python pymilvus_embedding.py   # Embed and insert data

Search and Evaluate

cd evaluation
python search_pymilvus.py      # Perform vector search
python evaluation.py           # Calculate Recall@k

🚧 TODO

• Add setup instructions for RAGatouille
• Complete evaluation for all datasets

---

📄 Citation

If you find this repository or its data useful, citing our paper would be appreciated:

ACL :

Hsing-Ping Liang, Che-Wei Chang, and Yao-Chung Fan. 2025. Improving Table Retrieval with Question Generation from Partial Tables. In Proceedings of the 4th Table Representation Learning Workshop, pages 217–228, Vienna, Austria. Association for Computational Linguistics.

Bibtex :

@inproceedings{liang-etal-2025-improving-table,
    title = "Improving Table Retrieval with Question Generation from Partial Tables",
    author = "Liang, Hsing-Ping  and
      Chang, Che-Wei  and
      Fan, Yao-Chung",
    booktitle = "Proceedings of the 4th Table Representation Learning Workshop",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.trl-1.19/",
    doi = "10.18653/v1/2025.trl-1.19",
    pages = "217--228",
}