This paper presents LLATMU's participation in the #SMM4H-HeaRD 2025 Task 3, which focused on identifying tweets that indicate caregiving for family members with dementia. We compare two approaches: a supervised BERTweet classifier and a generative Meta-LLaMA-3-8B model fine-tuned by Low-Rank Adaptation (LoRA). The BERTweet model achieved a higher F1-score (0.948) than the LLaMA-3 model (0.931), but the LLM model demonstrated strong contextual understanding, especially for vague or implicit caregiving references. By reformatting the task into a Yes/No prompt-based classification and incorporating a final verification layer using the OpenAI API, the LLaMA-3 system exhibited promising semantic inference capabilities with limited data. While traditional transformer models remain strong baselines, our findings suggest that fine-tuned LLMs are viable and scalable alternatives in low-resource settings. The use of parameter-efficient tuning and model quantization enabled the adaptation of a large LLM within constrained computational environments. Moreover, LLMs' multilingual capacity offers potential for broader public health surveillance. This study highlights the strengths and limitations of generative LLMs in health-related NLP tasks and points toward future directions in data augmentation, multilingual modeling, and real-time health signal detection.

PDF