Box-score data
A dataset of NBA game summaries aligned with box- and line-scores for data-to-text generation research.
What is Box-score data?
Boxscore Data is a research dataset that pairs NBA basketball game summaries with corresponding box-scores and line-scores. It was created to support data-to-document generation tasks, where models learn to produce human-like narratives from structured statistics. The dataset includes games from 2006 to 2017, sourced from Rotowire and SBNation, and is formatted in JSON for easy integration into machine learning pipelines.
Researchers and students in natural language processing, particularly those working on data-to-text generation, summarization, or sports analytics. It is also suitable for educators creating assignments on structured data processing.
This dataset provides a clean, aligned corpus of sports statistics and narratives, which is rare and valuable for training generative models. Its standardized format and preprocessed content reduce the overhead of data cleaning, allowing researchers to focus on model development and evaluation.
Overview
This dataset provides structured NBA basketball game data paired with human-written summaries, enabling research in data-to-document generation. It serves as a benchmark for training and evaluating models that convert structured statistics into coherent natural language narratives.
Key Features
- Aligned Summaries and Statistics — Each human-written game summary is paired with corresponding box-scores and line-scores.
- Dual Source Coverage — Includes data from Rotowire (2014–2017) and SBNation (2006–2017) with distinct writing styles.
- Structured JSON Format — Data is provided in a consistent JSON schema with team, player, and game details.
- Preprocessed for NLP — Summaries are tokenized and cleaned, with numeric values standardized as integers.
- Standard Splits — Data is divided into training, validation, and test sets for machine learning experiments.
Philosophy
The dataset is designed to support reproducible research in natural language generation, focusing on the challenge of transforming structured sports data into fluent, informative text.
Use Cases
Best For
- Training models for automatic sports article generation
- Benchmarking data-to-text generation systems
- Studying how structured data can be verbalized
- Educational projects in natural language processing
- Comparing narrative styles across different sports websites
- Research on multi-modal data integration (stats + text)
Not Ideal For
- Real-time sports analytics or live commentary systems requiring up-to-date data
- Projects needing multi-sport or non-basketball data for comparative analysis
- Applications that rely on raw, unprocessed web content without tokenization or cleaning
Pros & Cons
Pros
Each human-written game summary is directly paired with corresponding box-scores and line-scores, enabling straightforward training for data-to-text generation models as described in the README.
Includes data from Rotowire (2014–2017) and SBNation (2006–2017) with distinct writing styles, allowing researchers to compare narrative approaches and improve model robustness.
Summaries are tokenized with nltk, numbers standardized to integers, and irrelevant content like tweets removed, reducing preprocessing overhead for machine learning pipelines.
Data is provided in a consistent JSON schema with detailed team, player, and game objects, making it easy to parse and integrate into experimental setups.
Pre-divided into training, validation, and test sets for both Rotowire and SBNation data, facilitating reproducible research and benchmarking.
Cons
The README recommends using SportSett:Basketball instead due to contamination issues in the Rotowire dataset, where box- and line-scores appear in multiple splits, undermining data integrity for some experiments.
Covers NBA games only up to 2017, making it unsuitable for research requiring recent data or contemporary player and team performances.
Exclusively focuses on NBA basketball without inclusion of other sports or leagues, restricting its applicability to broader data-to-text tasks without significant adaptation.
Frequently Asked Questions
Related Projects
The Schema-Guided Dialogue DatasetThe Schema-Guided Dialogue Dataset
WebNLGThe enriched version of the WebNLG described at INLG 2018
Alex Context NLG DatasetDataset for NLG which contains preceding context along with each generation instance
Neural-WikipedianNeural-Wikipedian is a research project that adapts encoder-decoder neural network frameworks to automatically generate textual summaries (biographies) from structured Semantic Web triples. It addresses the challenge of transforming machine-readable knowledge base data into coherent, human-readable narratives, which is valuable for automating content creation and enhancing data accessibility. ## Key Features **Triple-to-Text Generation** — Converts sets of RDF triples (from DBpedia and Wikidata) into fluent English biography summaries. **Dual Dataset Support** — Includes aligned datasets of DBpedia and Wikidata triples paired with Wikipedia biographies for training and evaluation. **Neural Architectures** — Implements both Triples2LSTM and Triples2GRU models using the Torch framework for sequence generation. **Baseline Language Model** — Provides a KenLM n-gram language model as a comparative baseline for summary generation. **Pre-trained Models** — Offers downloadable pre-trained models for immediate inference without requiring full training cycles. ## Philosophy The project approaches biography generation as a structured data-to-text translation problem, leveraging neural networks to learn the linguistic patterns and factual associations present in Wikipedia content.
Found a gem we're missing?
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.