Motivation: Split-alignments provide base-pair-resolution evidence of genomic rearrangements. In practice, they are found by first computing high-scoring local alignments, parts of which are then combined into a split-alignment. This approach is challenging when aligning a short read to a large and repetitive reference, as it tends to produce many spurious local alignments leading to ambiguities in identifying the correct split-alignment. This problem is further exacerbated by the fact that rearrangements tend to occur in repeat-rich regions. Results: We propose a split-alignment technique that combats the issue of ambiguous alignments by combining information from probabilistic alignment with positional information from paired-end reads. We demonstrate that our method finds accurate split-alignments, and that this translates into improved performance of variant-calling tools that rely on split-alignments. Availability and implementation: An open-source implementation is freely available at: https://bitbucket.org/splitpairedend/last-split-pe. Supplementary information: Supplementary data are available at Bioinformatics online.