TREC-6 Text Classification Dataset with 10% Label Noise

Version of the TREC-6 question classification dataset with 10% artificially introduced label noise. The dataset is authored by rungalileo and was last updated in June 2022. The original TREC-6 dataset is used for classifying questions into six coarse-grained categories.

Use Cases

• Train robust text classifiers on noisy label data to improve model generalization.
• Benchmark noise-robust learning algorithms using the 10% label noise setting.
• Analyze the impact of label noise on classification performance across the six TREC question categories.

Strengths

• Based on the established TREC-6 benchmark for question classification.
• Provides a controlled setting with 10% label noise for robustness research.
• Last updated in 2022.

Limitations

• The specific number of rows, columns, and sample data are not provided.
• The nature and distribution of the introduced 10% label noise are not detailed.
• The dataset's size and file formats are unknown.

Provenance

Source

Derived from the TREC-6 dataset.

Collection Method

Original TREC-6 data with 10% of labels artificially altered to introduce noise.

Freshness

Last updated 2022-06-28.