By postulating closed-world
Assumption over
Ascent++, phrases which are
absent for the target concept
(pancake) are
assumed to be negative phrases.
Many are nonsensical.
By exploiting the chosen peers
and postulating
local closed-world assumption,
phrases become more thematic,
and
we see a
dramatic reduction in size (see
footer of the box). Crossed out
candidates in step-1 did not
make it to step-2 due to
irrelevance.
The candidates inferred in step-2
needs to be
scrutinized due to the high
incompleteness of the web-scale
knowledge
base. Crossed out candidates in
step-2 did not make it to step-3
and
are considered implausible
negations.
Not[[ asrt.predicate ]] → [[ asrt.object ]] ([[ sibling.list.length ]] [[ sibling.type ]]s: such as[[ s ]], ; )
The candidates inferred in step-3
are ranked by
salience, i.e., relative
frequency within the group of
peers. Read more
about the details in the
research
paper (Section 3.4).