Infimobius at #SMM4H-HeaRD 2026: Multi-Seed DeBERTa Ensemble for Flu Vaccination and Testing Status Classification

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

Summary

FluENS (Flu ENsemble System), a submission to the Social Media Mining for Health (SMM4H) 2026 Shared Task 3, performs fine-grained classification of flu vaccination and testing statuses from tweets. This system is built upon the "microsoft/deberta-v2-xlarge" pre-trained language model. It employs a multi-seed ensemble strategy, aggregating five models initialized with distinct random seeds and trained on the full dataset through soft-voting on averaged softmax probabilities. To mitigate severe label imbalance, FluENS incorporates balanced class weights and applies a two-stage learning rate schedule that independently controls the encoder and classification head. On the development set, FluENS achieved a macro F1 of 79.64% and micro F1 of 85.56% for the flu vaccination sub-task, and a macro F1 of 96.35% and micro F1 of 97.04% for the flu testing sub-task, surpassing a "roberta-base" baseline.

Key takeaway

For NLP Engineers developing social media health monitoring systems, consider adopting multi-seed ensemble strategies with models like "DeBERTa-v2-xlarge". Your classification performance for fine-grained tasks, such as flu vaccination status, can significantly improve by incorporating balanced class weights and a two-stage learning rate schedule. This approach yielded macro F1 scores up to 96.35% on flu testing, suggesting a robust method for enhancing public health surveillance accuracy.

Key insights

Multi-seed DeBERTa ensembles with specific training strategies significantly improve flu status classification from tweets.

Principles

• Ensemble methods enhance classification robustness.
• Addressing label imbalance improves model performance.
• Fine-grained learning rates optimize model components.

Method

Train five "microsoft/deberta-v2-xlarge" models with different random seeds on the full dataset. Aggregate via soft-voting on averaged softmax probabilities, using balanced class weights and a two-stage learning rate schedule.

In practice

• Implement multi-seed ensembling for robust NLP tasks.
• Apply balanced class weights for imbalanced datasets.
• Use two-stage learning rates for encoder/head optimization.

Topics

• Social Media Mining
• DeBERTa-v2-xlarge
• Ensemble Learning
• Flu Status Classification
• Natural Language Processing
• Health Informatics

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• MentalMARBERT: Domain-Adaptive Pre-training and Two-Stage Fine-Tuning for Arabic Mental Health Disorders Detection — Domain-adaptive pre-training and hierarchical classification significantly improve Arabic mental health disorder detection.
• Enhancing SDG-Text Classification with Combinatorial Fusion Analysis and Generative AI — Combinatorial Fusion Analysis (CFA) significantly enhances SDG text classification by combining diverse models and synthetic data.
• LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification — Iterative imbalance-aware fine-tuning of LLMs, combined with augmentation and post-processing, significantly improves rare-class NLP performance.
• AAbAAC: An Annotated Corpus for Autoimmunity Information Extraction — Small, specialized annotated corpora significantly boost information extraction performance in niche biomedical domains.
• Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection — Graph-based features can augment pun detection, but their utility depends on specific lexical relation encoding.
• Comparing BERT Sentence-Pair Classification and Few-Shot LLM Prompting for Detecting Threat and Solution Framing in German Climate News — Fine-tuned BERT models outperform few-shot LLMs for German climate news framing detection, especially with sentence context.

Open in AIssential →

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