neural abstractive summarization models are prone to generate content
inconsistent with the source document, i.e. unfaithful. Existing automatic
metrics do not capture such mistakes effectively. We tackle the problem of
evaluating faithfulness of a generated summary given its source do