DSGlobal

Overview

DSGlobal is the global-news counterpart to LocalNews, released with NEWSCOPE to test whether diverse-retrieval results generalize beyond regional reporting.

It is an adaptation, not a new collection: the underlying corpus is DiverseSumm (Huang et al., 2024), a dataset originally built for multi-source news summarization, re-segmented into paragraphs and repurposed as a retrieval benchmark. That lineage is the point — DiverseSumm was already assembled around the premise that one event is reported many ways, which is exactly the property a diversity benchmark needs.

Statistics

PropertyValue
Events147
Paragraphs7,532
Avg. sentences per paragraph7.5
Avg. words per paragraph123

Larger than LocalNews on both axes, with near-identical paragraph shape — the two were built to be compared directly.

What It Measures

The same six measures as LocalNews: precision, recall and F1 for relevance, plus Average Pairwise Distance, Positive Cluster Coverage and Information Density Ratio for diversity, at depths of 5, 10, 20 and 50.

Results on DSGlobal track LocalNews closely, which is what the dataset exists to establish. The one visible difference is that recall is much lower at shallow depths across every method — global events carry more relevant paragraphs, so the top 5 covers proportionally less of them.

Availability

Distributed as DSGlobal.zip in the NEWSCOPE repository. The repository carries no license file; the original DiverseSumm terms also apply upstream.

  • LocalNews — the companion local-news benchmark, built from scratch rather than adapted