New York Times Articles with Distant Relation Annotations

NYT-multi contains New York Times news articles from 1987-2007, distantly annotated with relations using FreeBase. The original dataset by Riedel et al. (2010) included 1.18 million sentences, later filtered by Zeng et al. (2018) to 66,195 sentences. DFKI-SLT hosts the current version on Hugging Face.

Use Cases

• Train models to extract relations between named entities from news sentences.
• Study the performance of distant supervision methods using FreeBase-annotated text.
• Analyze patterns in entity co-occurrence and relation types within a 20-year news corpus.
• Benchmark sentence classification models on the filtered set of 66,195 relation-bearing sentences.

Strengths

• Original corpus contains 1.18 million sentences, providing substantial textual data.
• Covers a 20-year time range (1987-2007) of news articles.
• Filtered version provides a focused set of 66,195 sentences with active relations.

Limitations

• Annotations are generated via distant supervision, which can introduce label noise.
• The filtered version excludes sentences over 100 words and those without an active relation, potentially removing context.
• Specific column names, data formats, and exact row counts for the current version are not detailed in the provided description.

Provenance

Source

New York Times news articles (1987-2007).

Collection Method

Distant annotation using FreeBase relations, followed by filtering for sentence length and active relations.

Time Range

1987 to 2007

Freshness

Dataset card was last updated on 2024-08 11.

Geography

United States (primary region for New York Times reporting).