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

Summary

The authors tackle SemEval-2026 Task 4, which asks systems to judge how similar fictional stories are in their narrative structure: theme, course of action, and outcome, rather than in wording. They avoid costly LLM reasoning at prediction time. LLMs are used only offline, and similarity itself comes from fine-tuned sentence-transformer embeddings.

For Track A, the system picks which of two candidates is closer to an anchor story. It fine-tunes all-mpnet-base-v2 with a triplet margin loss and freezes the lower layers. It also replaces named entities with consistent placeholders so the model cannot depend on overlapping names.

For Track B, which produces story embeddings, LLMs split each story into theme, plot, and outcome views. Each view gets its own projection head, trained contrastively, and an alignment loss pulls the fused vector toward the mean of the views.

The system ranked 14th of 42 teams in Track A and 6th of 25 in Track B. This suggests that compact, embedding-only models can stay competitive on narrative similarity.

Key points

• Track A: private score 0.6925 (top system: 0.7800). Setup:
  • all-mpnet-base-v2 producing 768-dim mean-pooled embeddings;
  • embeddings and the bottom 40% of transformer layers frozen;
  • triplet margin loss with cosine distance and m = 0.3.
• Pseudonymization: built with fastcoref and spaCy (en_core_web_trf), with an NER fallback for missed spans.
  • Entities become placeholders such as Character_A or Location_1.
  • One mapping is shared across all texts in a sample, so the same name always gets the same placeholder.
• Effect of pseudonymization on the Track A test set: it helped only the MPNet backbone, despite the paper describing the gains as consistent.
  • all-mpnet-base-v2: 0.6650 → 0.6925.
  • para-mul-mpnet-v2: 0.6625 → 0.6300.
  • LaBSE: 0.6050 → 0.5975.
  • all-MiniLM-L6-v2: 0.6125 → 0.5800.
• Track B: private score 0.6875 on Codabench (top system: 0.7200). gpt-4o-mini, gpt-4.1 and o3-mini were compared as extractors of the narrative views.
• Track B fusion weights (dev set): the hand-tuned weights (full 0.50, theme 0.10, plot 0.20, outcome 0.20) scored 0.725 on the dev set.
  • Full-text embedding only: 0.65.
  • Equal weights: 0.625.