Short reads can map break points
With short reads you do an "indirect observation/inference" of a structural variant: based on your paired-end reads you figure out something is wrong because their distance or orientation is not what you expect. With long reads, you often get a direct observation of a breakpoint, with sufficient long reads on both sides (either split alignment or spanning the SV) allowing almost nucleotide accuracy.
Long reads can disambiguate repeat related issues
Short reads cannot align unambiguously in many repetitive regions, because the repeat is longer than the read. Long reads, however, span most common repeats entirely, allowing you to unambiguously assign a location for each read, as such resolving the issues.