Model description

A BART (base) model trained to classify whether a summary is faithful to the original article. See our paper in NAACL'21 for details.


Concatenate a summary and a source document as input (note that the summary needs to be the first sentence).

Here's an example usage (with PyTorch)

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("CogComp/bart-faithful-summary-detector")
model = AutoModelForSequenceClassification.from_pretrained("CogComp/bart-faithful-summary-detector")

article = "Ban Ki-Moon was re-elected for a second term by the UN General Assembly, unopposed and unanimously, on 21 June 2011."

bad_summary = "Ban Ki-moon was elected for a second term in 2007."
good_summary = "Ban Ki-moon was elected for a second term in 2011."

bad_pair = tokenizer(text=bad_summary, text_pair=article, return_tensors='pt')
good_pair = tokenizer(text=good_summary, text_pair=article, return_tensors='pt')

bad_score = model(**bad_pair)
good_score = model(**good_pair)

print(good_score[0][:, 1] > bad_score[0][:, 1]) # True, label mapping: "0" -> "Hallucinated" "1" -> "Faithful"

BibTeX entry and citation info

    author = {Sihao Chen and Fan Zhang and Kazoo Sone and Dan Roth},
    title = {{Improving Faithfulness in Abstractive Summarization with Contrast Candidate Generation and Selection}},
    booktitle = {NAACL},
    year = {2021}
Downloads last month
Hosted inference API
Text Classification
Mask token: <mask>
This model can be loaded on the Inference API on-demand.