extractive (TextRank)
s1
s2
s3
s4
s5
s6
PageRank picks most connected sentences
verbatim. never hallucinates.
risk: misses distributed content.
abstractive (BART / T5 / Pegasus)
article (500-2000 tokens)
tokens in
transformer enc-dec
cross-attention
new text (60-200 tok)
generates new tokens, not selects
fluent, compressive.
risk: hallucinated entities,
flipped polarity, wrong numbers.