AG News Corpus with Over One Million Articles

A collection of over 1 million news articles collected from more than 2000 news sources over a period exceeding one year. It was compiled by the academic news search engine ComeToMyHead and adapted by Xiang Zhang for use as a text classification benchmark.

Use Cases

• Train character-level convolutional networks for text classification using the article text as demonstrated in the original NIPS 2015 paper.
• Benchmark clustering algorithms on a large-scale corpus of news articles from diverse sources.
• Conduct information retrieval research, such as ranking and search, on a collection of over 1 million documents.
• Analyze topic distribution and evolution across more than 2000 aggregated news sources.

Strengths

• Over 1 million individual news articles provides substantial volume for model training.
• Articles sourced from more than 2000 distinct news outlets offers diversity in content and perspective.
• Data collection spanned more than one year, capturing temporal variation in news topics.

Limitations

• Specific column structure, feature details, and label definitions for the classification benchmark are not provided in the input.
• The original corpus is over a decade old, which may limit relevance for modeling contemporary language or events.
• Potential geographic or linguistic bias is unknown as the source distribution of the 2000+ outlets is unspecified.

Provenance

Source

ComeToMyHead academic news search engine, adapted by Xiang Zhang.

Collection Method

Articles gathered via web crawling from more than 2000 news sources.

Time Range

Collection occurred over more than one year of activity, starting from July 2004.

Freshness

The dataset record was last updated on 2024-09-10, but the underlying article collection period concluded over a decade ago.