20 Newsgroups Text Collection for Document Classification

20,000 newsgroup documents organized into 20 different topics, originally collected by Ken Lang for machine learning research. The dataset is a standard benchmark for text classification and clustering algorithms. It is hosted on OpenML, a platform for sharing machine learning datasets and experiments.

Use Cases

• Classify document topic using the 20 predefined newsgroup labels as the target variable.
• Benchmark text vectorization methods (e.g., TF-IDF, word embeddings) on the document text features.
• Train multi-class classifiers like Naive Bayes or SVMs to predict the newsgroup from article content.
• Perform document clustering analysis to group similar articles without using the provided labels.
• Analyze term frequency and document length distributions across different thematic categories.

Strengths

• Approximately 20,000 total documents provides a substantial corpus for model training.
• Clear structure with 20 distinct, evenly distributed topics facilitates multi-class classification benchmarking.

Limitations

• Data is from the late 1990s, so vocabulary and topics may not reflect contemporary language or issues.
• Content is primarily from USENET discussion groups, which may introduce demographic and topical biases.

Provenance

Source

OpenML

Collection Method

Collected from USENET newsgroup postings in the 1990s.

Time Range

Circa 1990s