569 Polish news articles paired with 2,845 human-annotated extractive summaries and 770 abstractive summaries. Each article features five distinct extractive summaries representing approximately 5% of the source text, while a subset of 154 articles includes five additional abstractive summaries.
Use Cases
- Train extractive summarization models by predicting which sentences from the source text are included in the five human-selected summaries
- Benchmark abstractive summarization performance on Polish news text using the 770 human-written abstractive summaries
- Analyze annotator agreement and selection bias by comparing the five different extractive summaries provided for each article
Strengths
- 569 source news articles with five extractive summaries each
- Subset of 154 articles featuring five additional abstractive summaries
- Extractive summaries are constrained to approximately 5% of the original article length
- Part of the KLEJ (Kompleksowa Lista Ewaluacji Językowych) benchmark for Polish language processing