@inproceedings{bolisetty-etal-2026-moodmetric,
    title = "{M}ood{M}etric at {S}em{E}val-2026 Task 4:Narrative Story Similarity and Narrative Representation Learning",
    author = "Bolisetty, Samanvitha  and
      Ashar, Shreya  and
      Mittal, Nishchay  and
      Mishra, Pruthwik",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.336/",
    pages = "2664--2672",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper presents our system for narrative similarity modeling in SemEval Task 4, focusing on transformer-based dense embedding approaches. Modeling similarity between long-form narratives is particularly challenging due to the need to capture event progression, causal structure, character dynamics, and thematic coherence beyond surface-level lexical overlap.We evaluate multiple pretrained encoder-only architectures, including DeBERTa-v3, BGE-Base, BGE-Large, and E5-Large, fine-tuned using triplet margin and contrastive objectives. In addition, we implement a hybrid lexical{--}semantic baseline combining TF-IDF and SBERT features. Our experiments analyze the impact of model scale, pooling strategies, layer freezing, training duration, and embedding-level ensembling under low-resource conditions (approximately 1,900 training triplets, with additional synthetic augmentation).Results show that larger contrastively pretrained embedding models consistently outperform smaller variants, with BGE-Large achieving the strongest standalone performance. However, performance saturates quickly, and moderate fine-tuning (4{--}5 epochs) yields optimal validation accuracy, while extended training leads to overfitting. Instruction-tuned embeddings do not demonstrate significant advantages over contrastively aligned alternatives for this task. Finally, arithmetic averaging of embeddings from diverse models produces the most robust representations, achieving approximately 65{\%} validation accuracy."
}