@article{van-nooten-kosar-2026-push,
    title = "Push and Pull: Training Sentence Encoders with Contrastive Losses for Distance-Based Multi-Label Text Classification",
    author = "Van Nooten, Jens  and
      Kosar, Andriy",
    editor = "Piperidis, Stelios  and
      Bel, N{\'u}ria  and
      van den Heuvel, Henk  and
      Ide, Nancy  and
      Krek, Simon  and
      Toral, Antonio",
    journal = "International Conference on Language Resources and Evaluation",
    volume = "main",
    month = may,
    year = "2026",
    address = "Palma de Mallorca, Spain",
    publisher = "ELRA Language Resource Association",
    url = "https://preview.aclanthology.org/ingest-lrec/2026.lrec-main.583/",
    pages = "7359--7379",
    abstract = "Despite the potential of Distance-Based Classification (DBC), a method that assigns labels to text by measuring semantic similarity between the text and the label representations, it has received very little attention for Multi-Label Text Classification (MLTC). Previous studies have focused on determining optimal thresholds, reaching promising results with contextual sentence encoders. We demonstrate that the performance of these models can be further improved by training them with contrastive losses, i.e., by bringing text representations closer to the corresponding true label representations in an embedding space. Using three supervised contrastive losses and three sentence encoders (Stella, GIST-Large, and BGE), we evaluated our approach on five English datasets (SemEval, BioTech, Reuters, AAPD, and LitCovid) and one Dutch dataset (EventDNA). The results show consistent substantial improvements over base sentence encoders, thereby narrowing the gap between DBC methods and fine-tuned or zero-shot approaches."
}