This repository contains the faithfulness and factuality annotations of XSum summaries from our paper "On Faithfulness and Factuality in Abstractive Summarization" at ACL 2020.
Please cite our paper if you use our data.
@InProceedings{maynez_acl20,
author = "Joshua Maynez and Shashi Narayan and Bernd Bohnet and Ryan Thomas Mcdonald",
title = "On Faithfulness and Factuality in Abstractive Summarization",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
year = "2020",
pages = "1906--1919",
address = "Online",
}
Neural abstractive summarization models are highly prone to hallucinate content that is unfaithful to the input document. The popular metric such as ROUGE fails to show the severity of the problem. Our dataset is unique in that we have conducted a large scale human evaluation of several neural abstractive summarization systems to better understand the types of hallucinations they produce. Our human annotators found substantial amounts of hallucinated content in all model generated summaries. However, our analysis does show that pretrained models are better summarizers not only in terms of raw metrics, i.e., ROUGE, but also in generating faithful and factual summaries as evaluated by humans. Furthermore, we show that textual entailment measures better correlate with faithfulness than standard metrics, potentially leading the way to automatic evaluation metrics as well as training and decoding criteria. We believe that this dataset will help advance the quality of neural summarization by enabling research to better train and evaluate for trustworthy systems that could have been more challenging to do otherwise.
The dataset consists of faithfulness and factuality annotations of abstractive summaries for the XSum dataset. We have crowdsourced 3 judgements for each of 500 x 5 document-system pairs. This will be a valuable resource to the abstractive summarization community.
Raters are shown the news article and the system summary, and are tasked with identifying and annotating the spans that aren't supported by the input article. The csv file contains the following columns:
bbcid: Document id in the XSum corpus.
system: Name of neural summarizer.
summary: Summary generated by ‘system’.
hallucination_type: Type of hallucination (intrinsic or extrinsic)
hallucinated_span: Hallucinated span in the ‘summary’.
hallucinated_span_start: Index of the start of the hallucinated span.
hallucinated_span_end: Index of the end of the hallucinated span.
worker_id: 'wid_0', 'wid_1', 'wid_2'
Raters are shown the news article and the hallucinated system summary, and are tasked with assessing the summary whether it is factual or not. The csv file contains the following columns:
bbcid: Document id in the XSum corpus.
system: Name of neural summarizer.
summary: Summary generated by ‘system’.
is_factual: yes/no
worker_id: 'wid_0', 'wid_1', 'wid_2'
Precomputed ROUGE, BERTScore, entailment faithfulness and factuality scores for each system and BBC document pair. For faithfulness score, we map the hallucination spans for each summary to word level. We assign a score of 1.0 to each word if it is not in one of the hallucination spans marked by an annotator. Finally we take the average over the number of annotations (3) and the number of words in the summary to get the final faithfulness score for each summary. For factuality score, we assign a score of 1.0 to a summary when annotated factual and 0.0 when annotated not-factual. We take the average of all three annotation scores to get the final factuality score for each summary.
system_bbcid: System id and BBC document id.
R1/R2/RL: ROUGE F1 scores.
BERTScore: BERTScore (Zhang et al. 2019).
Entailment: Entailment probability.
Faithful: Faithfulness score.
Factual: Factuality score.
This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0).
If you have a technical question regarding the dataset or publication, please create an issue in this repository. This is the fastest way to reach us.
If you would like to share feedback or report concerns, please email us at xsum-hallucinations-acl20@google.com.