Table 3:

Methods to create negative and positive candidates in support of relevance and faithfulness calibration, respectively. For each candidate generation method, we include whether it is used as a positive or negative example (both in the case of relevance ranking), what inputs it requires (the source document and/or the reference (ref.)), as well as the external components needed and, finally, the specific models used for the experiments in this paper.

	Method	−	+	Source	Ref.	External Components	Models Used
Relevance Calibration	Diverse Beam	✓	✓	✓		Summarization Model	PRIMERA
	Diverse Beam	✓	✓	✓		Summarization Model	LongT5
Faithful Calibration	Mask-And-Fill	✓			✓	Constituency Parser, PLM	Stanza, SciFive
	Entity Swap	✓			✓	Entity, Number Extractors	BERN2, Quantulum
	Paraphrase		✓		✓	Paraphrase Generator	GPT-3 + Curated Prompt
	Reference		✓		✓	N/A	N/A