Team Paradise at #SMM4H-HeaRD 2026: Multi-Task Approaches for Social Media Health Mining

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Public Health & Epidemiology · Depth: Expert, quick

Summary

Team Paradise presented its systems for three tasks in the SMM4H-HeaRD 2026 shared task, focusing on multi-task approaches for social media health mining. For multilingual adverse drug event detection (Task 1), their XLMRoBERTa-large model, with threshold-only ablation, achieved a macro-F1 of 0.597, surpassing the field mean of 0.547 by +0.050. In Task 3, influenza vaccine effectiveness estimation, a three-stage hybrid pipeline combining twitter-RoBERTa-base-2022 with rule-based post-processing yielded Micro-F1 scores of 0.8434 for vaccination status and 0.8936 for test results. For opioid impact span extraction (Task 7), a RoBERTa-large model utilizing CRF decoding and sliding-window inference obtained a relaxed F1 of 0.60, despite significant train-test distributional shift. Across all tasks, the team identified class imbalance, temporal ambiguity, and platform heterogeneity as key challenges.

Key takeaway

For NLP Engineers developing social media health mining systems, consider integrating multi-task approaches to improve performance across diverse health-related classifications and extractions. Your models, like XLMRoBERTa-large or RoBERTa-large, can benefit from techniques such as threshold-only ablation for classification or CRF decoding with sliding-window inference for span extraction. Be prepared to address challenges like class imbalance and platform heterogeneity to maintain robust system performance.

Key insights

Multi-task approaches using large language models effectively address diverse social media health mining challenges.

Principles

• Threshold-only ablation can boost F1 scores.
• Hybrid pipelines improve classification accuracy.
• CRF decoding aids span extraction.

Method

A three-stage hybrid pipeline combines a RoBERTa-base model with rule-based post-processing for classification, while CRF decoding and sliding-window inference are used for span extraction.

In practice

• Apply XLMRoBERTa-large with threshold ablation.
• Combine RoBERTa with rule-based post-processing.
• Use CRF decoding for sequence labeling tasks.

Topics

• Social Media Mining
• Health Research
• Adverse Drug Events
• Vaccine Effectiveness
• Opioid Impact
• RoBERTa Models
• Natural Language Processing

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• ChatHealthAI: Aligning Electronic Health Record Representations with Large Language Models for Grounded Clinical Reasoning — ChatHealthAI integrates EHR representations with LLMs for grounded, interpretable clinical reasoning and accurate patient prediction.
• Gated Multi-Graph Fusion via Graph Attention Networks for Alzheimer's Disease Detection — Gated Multi-Graph Fusion via Graph Attention Networks improves Alzheimer's Disease detection from speech by integrating diverse linguistic features.
• AI scans 400,000 Reddit posts and finds hidden Ozempic side effects — AI-driven social media analysis can rapidly identify underreported drug side effects missed by traditional clinical trials.
• RecallRisk-BERT: A Multi-Task Framework for Post-Report Medical Device Recall Triage — Multi-task text-tabular models effectively triage medical device recalls by jointly predicting severity and root-cause.
• When PCOS Meets Eating Disorders: An Explainable AI Approach to Detecting the Hidden Triple Burden — Explainable AI can detect co-occurring health conditions like PCOS and eating disorders from social media data.
• 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.

Open in AIssential →

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