@inproceedings{pagel-reiter-2026-spinfo,
    title = "Spinfo Cologne at {S}em{E}val-2026 Task 4: Explainable Creation of Narrativity Embeddings",
    author = "Pagel, Janis  and
      Reiter, Nils",
    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-1.299/",
    pages = "2376--2383",
    ISBN = "979-8-89176-414-9",
    abstract = "We describe our submission to SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning.The task requires (i) selecting, for a given anchor story summary, which of two candidate summaries is narratively closer (Track A) and (ii) producing a narrative representation of a story as a vector embedding (Track B).Our approach emphasizes interpretability by explicitly eliciting three narrativity aspects with a prompted large language model.We then construct a fixed-size narrative embedding by concatenating aspect-wise representations, comparing a static-embedding baseline (GloVe) to contextualized sentence-transformer embeddings (all-MiniLM-L6-v2).On the development set, the sentence-transformer variant outperforms the static baseline and achieves 61.5{\%} accuracy on Track A, while the GloVe variant performs near chance.Our official submission reaches 60.25{\%} accuracy on the Track A test set and 57.75{\%} accuracy on Track B.Additional ablations show that the aspect pipeline slightly outperforms raw-text embeddings, but that aspect contributions are uneven.Qualitative analysis suggests that failures often stem from inconsistent aspect generation and from overemphasizing theme overlap over event-level similarity."
}