USB: A Unified Summarization Benchmark Across Tasks and Domains

This benchmark contains labeled datasets for 8 text summarization based tasks given below. The labeled datasets are created by collecting manual annotations on top of Wikipedia articles from 6 different domains.

Task	Description	Code snippet
Extractive Summarization	Highlight important sentences in the source article	load_dataset("kundank/usb","extractive_summarization")
Abstractive Summarization	Generate a summary of the source	load_dataset("kundank/usb","abstractive_summarization")
Topic-based Summarization	Generate a summary of the source focusing on the given topic	load_dataset("kundank/usb","topicbased_summarization")
Multi-sentence Compression	Compress selected sentences into a one-line summary	load_dataset("kundank/usb","multisentence_compression")
Evidence Extraction	Surface evidence from the source for a summary sentence	load_dataset("kundank/usb","evidence_extraction")
Factuality Classification	Predict the factual accuracy of a summary sentence with respect to provided evidence	load_dataset("kundank/usb","factuality_classification")
Unsupported Span Prediction	Identify spans in a summary sentence which are not substantiated by the provided evidence	load_dataset("kundank/usb","unsupported_span_prediction")
Fixing Factuality	Rewrite a summary sentence to remove any factual errors or unsupported claims, with respect to provided evidence	load_dataset("kundank/usb","fixing_factuality")

Additionally, to load the full set of collected annotations which were leveraged to make the labeled datasets for above tasks, use the command: load_dataset("kundank/usb","all_annotations")

Trained models

We fine-tuned Flan-T5-XL models on the training set of each task in the benchmark. They are available at the links given below:

Task	Finetuned Flan-T5-XL model
Extractive Summarization	link
Abstractive Summarization	link
Topic-based Summarization	link
Multi-sentence Compression	link
Evidence Extraction	link
Factuality Classification	link
Unsupported Span Prediction	link
Fixing Factuality	link

More details can be found in the paper: https://aclanthology.org/2023.findings-emnlp.592/

If you use this dataset, please cite it as below:

@inproceedings{krishna-etal-2023-usb,
    title = "{USB}: A Unified Summarization Benchmark Across Tasks and Domains",
    author = "Krishna, Kundan  and
      Gupta, Prakhar  and
      Ramprasad, Sanjana  and
      Wallace, Byron  and
      Bigham, Jeffrey  and
      Lipton, Zachary",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
    year = "2023",
    pages = "8826--8845"
}