hermeneutichools at SemEval-2026 Task 4: Multiperspectivity as a Resource for Narrative Similarity Prediction

AbstractPredicting narrative similarity can be under-stood as an inherently interpretive task: differ-ent, equally valid readings of the same text canproduce divergent interpretations and thus dif-ferent similarity judgments, posing a fundamen-tal challenge for semantic evaluation bench-marks that encode a single ground truth. Ratherthan treating this multiperspectivity as a chal-lenge to overcome, we propose to incorporateit in the decision making process of predic-tive systems. To explore this strategy, we cre-ated an ensemble of 31 LLM personas. Theserange from practitioners following interpretiveframeworks to more intuitive, lay-style charac-ters. Our experiments were conducted on theSemEval-2026 Task 4 dataset, where the sys-tem ranked 13th out of 47 teams and achievedan accuracy score of 0.705. Accuracy improveswith ensemble size, consistent with CondorcetJury Theorem-like dynamics under weakenedindependence. Practitioner personas performworse individually but produce less correlatederrors, yielding larger ensemble gains undermajority voting. Our error analysis reveals aconsistent negative association between gender-focused interpretive vocabulary and accuracyacross all persona categories, suggesting ei-ther attention to dimensions not relevant for thebenchmark or valid interpretations absent fromthe ground truth. This finding underscores theneed for evaluation frameworks that accountfor interpretive plurality.