IITPatna_ADE at #SMM4H-HeaRD 2026: Multilingual Adverse Drug Event Detection with LoRA-XLM-RoBERTa, Cross-Fold Ensembles, and Post-hoc Calibration

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Health & Wellbeing — Pharmaceuticals & Biotechnology, Public Health & Epidemiology, Medical Devices & Health Technology · Depth: Expert, short

Summary

The IITPatna_ADE system, submitted to Task 1 of #SMM4H-HeaRD 2026, addresses multilingual binary classification of adverse drug event (ADE) mentions in social media. This system fine-tunes "xlm-roberta-large" using LoRA adapters and learned language embeddings. Its training involves a two-stage process: initial domain adaptation with translated CADEC data, followed by five-fold cross-validation on the official training set. The approach further incorporates an ensemble of the five fold checkpoints via mean logits, applies temperature scaling on the development set, and tunes decision thresholds to optimize the official metric. On the development set, the ensemble achieved a macro-F1 of 0.788 with a global threshold and 0.796 with per-language thresholds. The best official test submission recorded a macro-F1 of 0.616 (ID 678990).

Key takeaway

For NLP Engineers developing multilingual health information extraction systems, this work demonstrates a robust methodology for adverse drug event detection. You should consider integrating LoRA-based fine-tuning with "xlm-roberta-large" for efficiency and performance. Implementing a two-stage training approach, combining domain adaptation with cross-validation, alongside ensemble methods and post-hoc calibration, can significantly improve macro-F1 scores in similar social media text classification tasks.

Key insights

Multilingual ADE detection benefits from LoRA-XLM-RoBERTa, two-stage training, cross-fold ensembles, and post-hoc calibration.

Principles

• Two-stage training improves domain adaptation.
• Ensembling cross-validation checkpoints enhances robustness.
• Post-hoc calibration refines model output for specific metrics.

Method

Fine-tune "xlm-roberta-large" with LoRA and language embeddings. Use two-stage training (CADEC domain adaptation, then 5-fold CV). Ensemble checkpoints by mean logits, apply temperature scaling, and tune decision thresholds.

In practice

• Implement LoRA for efficient fine-tuning of large models.
• Utilize cross-validation ensembles for robust predictions.
• Apply temperature scaling and threshold tuning for metric optimization.

Topics

• Multilingual NLP
• Adverse Drug Event Detection
• LoRA
• XLM-RoBERTa
• Ensemble Learning
• Social Media Mining
• Post-hoc Calibration

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Automated Detection of Dosing Errors in Clinical Trial Narratives: A Multi-Modal Feature Engineering Approach with LightGBM — Combining sparse lexical features with dense semantic embeddings improves clinical dosing error detection.
• System Report for CCL25-Eval Task 5: New Dataset and LoRA-Fine-Tuned Qwen2.5 — Domain-specific LoRA fine-tuning on Qwen2.5-14B with CCPoetry-49K significantly enhances classical Chinese poetry understanding and translation.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.
• Anatomy of Data Repositories for the Analysis and Detection of Toxicity in Portuguese — Lexical diversity and annotation practices significantly impact cross-domain generalization in hate speech detection datasets.
• CURA: Clinical Uncertainty Risk Alignment for Language Model-Based Risk Prediction — CURA improves clinical LM risk prediction by aligning uncertainty with individual errors and cohort ambiguities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.