@inproceedings{pachori-2026-bhramastra,
    title = "Bhramastra at {\#}{SMM}4{H}-{H}ea{RD} 2026: A Multi-Stage Hunter-Judge Pipeline using {DSP}y-Optimized {LLM}s for Multilingual {ADE} Detection",
    author = "Pachori, Bhaarat",
    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-1.9/",
    pages = "49--55",
    ISBN = "979-8-89176-432-3",
    abstract = "This paper describes the submission by **Team Bhramastra** for the **{\#}SMM4H-HeaRD 2026** Shared Task 1, focused on personal Adverse Drug Event (ADE) detection in multilingual social media. A decoupled architecture, **Hunter-Judge**, is proposed to handle extreme class imbalance and linguistic variance across seven languages, including a surprise zero-shot Farsi set. The system employs a fine-tuned multilingual mDeBERTa-v3 model as a high-recall filter (**Hunter**), followed by a Gemini-2.5-Flash model (**Judge**) optimized via the **DSPy** framework for precision-oriented agentic adjudication. By implementing a reasoning protocol grounded in clinical RAG evidence and universal ingredient mapping, the pipeline achieved the **highest average F1-score (0.6653)** among all teams. Strong zero-shot generalizability on Farsi (**F1: 0.5863**) was demonstrated, highlighting the effectiveness of medically-grounded adjudication in low-resource contexts."
}