(2025) - Zain Muhammad Mujahid, Dustin Wright, Isabelle Augenstein
Notes
Was recommended this paper as they are doing dataset pertubation, which what I am also interested in doing — pertubation → see effects in terms of preference.
Main metric here is factuality score, which is the log-likelihood scores. Figure 2 is the change in factuality score (perturbed minus original). In the figure, there are obvious shifts. -ve shift means higher score for pertubed. +ve shift means higher original score.
Pertubations used:
- Paraphrased — summary is rewritten with alternate phrasings and syntactic structures
- Simplified — where complex or compound constructions are rewritten into shorter, more readable sentences
- Synonym replaced — where content words are substituted with close synonyms to test for lexical invariance
- Less diverse — summaries that reduce vocabulary variation
- Negated — introduce logically equivalent negations to prbe sensititvity to syntactic polarity
- Summarized — further compresses the summary for brevity
- Added source text — inserts a factual sentence directly from the source
Another finding is that claim that are more general (highly similar to many other claims in the original document) are harder to fact-check.
- Measured by mean pair-wise cosine similarity between each summary sentence and the entire document.
- Higher cosine sim → more general
Thoughts
The pertubation methods should be kept in mind. But running this entire thing in a big SFT dataset seems impossible.