DUTIR at SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning

AbstractThis paper presents our approach for SemEval 2026 Task 4. Our method leverages a large language model fine-tuned via Low-Rank Adaptation, incorporates data cleaning, and employs a multi-prompt strategy, all trained on the official synthetic dataset. Evaluated on Track A, our system achieved an official score of 0.70, representing a reasonable performance under the given task constraints. In addition, we explore an alternative contrastive learning framework originally designed for Track B, where narrative-structure embeddings are learned and subsequently applied to Track A via similarity comparisons. Our analysis suggests that direct supervised adaptation may be more suitable for narrative reasoning tasks.