@article{rammohan-etal-2026-leveraging,
    title = "Leveraging Semi-Supervised Learning for Multimodal Hate Speech Data Annotation and Detection",
    author = {Rammohan, Rathi Adarshi  and
      Ren, Zhao  and
      Pucha{\l}a, Dominik  and
      {\'S}widerska, Aleksandra  and
      K{\"u}ster, Dennis  and
      Schultz, Tanja},
    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.806/",
    pages = "10266--10275",
    abstract = "While the Internet and social media have fundamentally transformed our lives, they can also rapidly spread hate speech, i.e., derogatory statements targeting individuals or groups based on their immutable characteristics. Automatic detection systems could help limit this harmful phenomenon. However, the lack of large-scale annotated datasets remains a major bottleneck for developing better algorithms. In this work, we employ semi-supervised learning (SSL) to leverage the advantages of limited labeled data alongside large amounts of unlabeled data. We apply three SSL approaches, Fix-match, Full-match, and All-match learning, to enhance the performance of end-to-end pre-trained speech and text models for hate speech detection. Our findings indicate that SSL methods enhance the performance, achieving F1 scores of 0.851 on speech, 0.957 on text, and 0.959 with multimodal fusion. Furthermore, we analyze the impact of different weak augmentation strategies on labeled data and assess the quality of generated pseudo-labels to evaluate their potential use in data annotation."
}