@inproceedings{danu-2026-siemens,
    title = "{SIEMENS} at {\#}{SMM}4{H}{--}{H}ea{RD} 2026: The Impact of Training Strategy and Backbone Selection on {BERT}-based Multilingual Clinical {NER}",
    author = "Danu, Manuela Daniela",
    editor = "Lopez-Garcia, Guillermo  and
      Gonzalez-Hernandez, Graciela",
    booktitle = "Proceedings of the 11th Social Media Mining for Health Research and Applications ({SMM}4{H}-{H}ea{RD} 2026) Workshop and Shared Tasks",
    month = jul,
    year = "2026",
    address = "San Diego, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.smm4h-1.34/",
    pages = "216--221",
    ISBN = "979-8-89176-432-3",
    abstract = "This paper describes our participation in the MultiClinNER subtask of the MultiClinAI shared task, part of the {\#}SMM4H-HeaRD Workshop at ACL 2026. The task requires identifying DISEASE, SYMPTOM, and PROCEDURE mentions in clinical case reports across seven languages: Czech, Dutch, English, Italian, Romanian, Spanish, and Swedish. We compare two BERT-based sequence labeling methods: (i) sentence-level token classification with a fixed train/validation split, and (ii) paragraph-level chunking with 5-fold cross-validation and checkpoint merging, using language-specific BERT models and multilingual XLM-RoBERTa-large as backbones. Our results show that 5-fold training with checkpoint merging consistently outperforms the fixed split strategy, with further analysis suggesting that the gains are primarily driven by improved training-set coverage rather than by differences in input granularity. Language-specific BERT encoders prove most effective for Spanish and English, while XLM-RoBERTa-large yields the strongest results for the remaining five languages through cross-lingual transfer."
}