📄 QASports2: Question-Answering Dataset about Sports

The first large-scale, open-domain sports question answering dataset

QASports is a comprehensive dataset featuring over 1 million question-answer-context tuples derived from more than 400,000 thoroughly preprocessed, cleaned, and organized documents about players, teams, and matches from multiple sports. The data is sourced from Wikipedia-like resources to ensure quality and relevance.

📚 Research Paper

Abstract

Sport is an engaging topic that is quickly evolving due to its growing popularity and revenue potential. This theme introduces several opportunities for question-answering (QA) systems, such as supporting tactical decision-making. Nevertheless, these QA systems require expert, specialized datasets and models. To advance the field of sports question-answering, we firstly present QASports2, a novel and significantly expanded dataset. QASports2 compiles information from 20 different wiki sports sources, resulting in over 1 million context-question-answer tuples. Subsequently, we introduce a data labeling validation algorithm that leverages both cluster sampling and large language models (LLMs). To corroborate the LLM results, we conducted a detailed manual review to ensure the labeling was accurate. We found Qwen2 to have the closest LLM results when compared to human labeling, reaching 83.9% question and 87.5% answer accuracy. Additionally, we conduct 399 comparative analysis of language models on QASports2 to evaluate their effectiveness as document reader and retriever models. We highlight the best models, BM25 and MiniLM which obtained 90.8% recall@20 and 93.4% f-score, respectively. Besides, this work outlines different aspects and applications this real-world sports data.

🚀 Quick Start

Prerequisites

• Python 3.10+
• uv package manager (Installation Guide)

Installation

```bash

Clone the repository

git clone https://github.com/leomaurodesenv/qasports-dataset-scripts.git cd qasports-dataset-scripts

Install dependencies

uv sync

Verify installation

uv run pre-commit run --all-files ```

📥 Download the Dataset

• 🎲 Full Dataset: OSF Repository
• 🎲 Formatted Dataset: Hugging Face Hub
• 🛠 Dataset v1: GitHub Release v1.1.0

🏗️ Dataset Generation Pipeline

The dataset generation process is organized into seven main stages, each contained in separate modules:

📁 Project Structure

src/ ├── crawler/ # 🔍 Gather wiki links ├── fetching/ # 📥 Fetch raw HTML from links ├── processing/ # 🧹 Process and clean textual data ├── extracting_context/ # 📄 Extract contexts from data ├── question_answer/ # ❓ Generate questions and answers ├── sampling/ # 🎯 Sample representative questions └── labeling_llm/ # 🏷️ Label samples using LLMs

🔄 Generation Steps

```bash

1. Crawler: Gather wiki links (~2 minutes)

uv run -m src.crawler.run

2. Fetching: Download wiki pages (~20 hours)

uv run -m src.fetching.run

3. Processing: Clean and process text (~50 minutes)

uv run -m src.processing.run

4. Context Extraction: Extract relevant contexts (~35 seconds)

uv run -m src.extracting_context.run

5. Q&A Generation: Create questions and answers (~5 days)

uv run -m src.questionanswer.run uv run -m src.questionanswer.run_huggingface # Optional

6. Sampling: Select representative questions (~1h 30 minutes)

uv run -m src.sampling.run

7. LLM Labeling: Label samples using LLMs (~91 hours)

uv run -m src.labeling_llm.run ```

🧪 Experiments

This repository includes experimental frameworks for evaluating QA systems using the QASports dataset.

Document Retriever Experiments

```bash

Run document retriever experiments (~24 days)

uv run -m experiments.doc_retriever --help

Example usage

uv run -m experiments.docretriever --model BM25 --numk 3 ```

Document Reader Experiments

```bash

Run document reader experiments (~37 hours)

uv run -m experiments.doc_reader --help

Example usage

uv run -m experiments.doc_reader --model RoBERTa --dataset SQuAD ```

📊 Dataset Statistics

• Total Questions: 1,000,000+
• Source Documents: 400,000+ preprocessed documents
• Data Sources: Wikipedia-like resources
• Question Types: Extractive QA, Wh-questions
• Sports Covered: Football, American Football, Basketball, Cricket, +15 Sports

```bash

Run dataset general analysis

uv run -m experiments.dataset_analysis --help

Example usage

uv run -m experiments.dataset_analysis --dataset QASports --sport RUGBY ```

🤝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Brazilian Computer Society for hosting the Dataset Showcase Workshop
• The research community for feedback and contributions
• All contributors to the QASports dataset

📖 Citation

If you use QASports in your research, please cite our paper:

bibtex @inproceedings{jardim:2023:qasports-dataset, author={Pedro Calciolari Jardim and Leonardo Mauro Pereira Moraes and Cristina Dutra Aguiar}, title = {{QASports}: A Question Answering Dataset about Sports}, booktitle = {Proceedings of the Brazilian Symposium on Databases: Dataset Showcase Workshop}, address = {Belo Horizonte, MG, Brazil}, url = {https://github.com/leomaurodesenv/qasports-dataset-scripts}, publisher = {Brazilian Computer Society}, pages = {1-12}, year = {2023} }