ttda704 at SemEval-2026 Task 4: Modeling Narrative Structures via Pseudonymization and Multi-View Sentence Alignment

The article presents the approach of team ttda704 for SemEval 2026 Task 4, focusing on narrative story similarity and representation learning. They propose two pipelines: a single-view method using fine-tuned sentence transformers with layer freezing to mitigate overfitting, and a multi-view method that incorporates view-specific projection heads for modeling themes, plots, and outcomes through self-supervised alignment. This work emphasizes the use of contrastive learning on synthetic data, which could enhance practitioners' capabilities in narrative analysis and similarity tasks using LLMs.