… examples of generated NLI instances in Table 6. We also include cases where Falsesum
inadvertently generates factually … several examples of the formatted input and the generated …

… models often generate summaries that are inconsistent with … tion, a document-level Natural
Language Inference (NLI) … two standard benchmarks, FALSESUM and SUMMAC. Compared …

… of factual inconsistencies in generated sentences by identifying … For example, given an NLI
model M(p, h), source document … -edge NLI based factual inconsistency detection method …

… • Finally, we present a document-level factuality classifier … train a factual consistency
classifier using an NLI model as … Realizing the diversity of inconsistencies and the challenges …

… sentence is inconsistent with respect to the document. Entity Error (… The success of the
documentlevel approach lies on the fact that … We gratefully acknowledge the support of the UK …

… the task of localizing factual inconsistencies as identifying the set of “… Supervised We apply
two recent off-the-shelf NLI models to … For example, the sentences “John missed the train and …

… derive stronger Natural Language Inference (NLI) models, we … of generating a binary decision
directly. A summary is thus … 50k examples from DocNLI and FalseSum respectively that do …

… examples that are diverse and inconsistent yet plausible. We show that models trained on a
Falsesum-augmented NLI … benchmarks for detecting factual inconsistency in summarization. …

… are many factual inconsistencies in the model-generated … levels of granularity, such as
document-level and sentence-level. [… Constructing training dataset for classification models in NLI-…

… framework that generates factually inconsistent summaries … for identifying the factual
inconsistencies within generated … to the data generated by FACTCC and FALSESUM and re-train …