CophiWue at SemEval-2026 Task 4: Symbolic Narrative Profiling with Taxonomy-Guided Extraction and Contrastive Fine-Tuning

AbstractWe present our system for SemEval-2026 Task 4, focusing primarily on Track B (narrative embedding). Our approach, the Decompose & Align Cycle, converts each story into a structured NarrativeProfile consisting of abstract themes, a five-step course of action, and an outcome. We then build a NarrativeTaxonomy from these initial extractions via agglomerative clustering, and use the resulting controlled vocabularies to guide a second extraction pass, producing terminologically standardized profiles across the full dataset. Finally, we contrastively fine-tune the Qwen3-Embedding8B model on profile text representations using TripletLoss, deriving story embeddings from this fine-tuned model. For Track A, we adapt the task’s provided baseline script by substituting Gemini 3 Pro as the judge, using the organizers default prompt on raw story texts.