Cross-lingual transfer with multilingual language models for influenza-like-illness detection in social media texts

Current Language Models (LMs) can process multilingual texts, enabling social media-based Influenza-Like-Illness (ILI) surveillance across languages without requiring language-specific systems. Yet, it remains unclear whether a cost-effective multilingual classifier can reliably extract ILI signals from social media across countries, languages, and seasons. We introduce MuILI, the first class-balanced multilingual dataset annotated for ILI detection, comprising 4,284 tweets across five widely spoken European languages. Using MuILI, we investigated whether a single multilingual classifier can consistently extract ILI signals across nine ILI seasons for France (2014-2023) and five seasons for Italy, Spain, and Germany (2018-2023). We compared five fine-tuned Multilingual-Masked LMs (MMLMs), three Cross-Lingual Transfer (CLT) strategies, and four In-Context Learning (ICL) techniques with decoder-only Large LMs (LLMs). Domain-adapted MMLMs (Bernice, F1 = 0.87), set the benchmark performance. Comparable results were obtained with translate-train (F1 = 0.85) or model-transfer (F1 = 0.84) from French data alone. Spanish translate-train performed similarly (F1 = 0.84). Among LLMs, only Qwen2.5-72B-Instruct surpassed the benchmark (F1 = 0.88-0.90). Spearman cross-correlation between ILI signals derived from the benchmark classifier and official consultation rates demonstrated strong real-time alignment, though it varied during disruptions. Overall, multilingual LMs can expand ILI surveillance using social-media. CLT offers annotation-efficient solutions, while ICL with LLMs can eliminate annotation requirements but at substantial computational costs.