Detecting vaccine adverse events in noisy social media texts is both challenging and crucial to timely pharmacovigilance. In this paper, we detail our participation in Task 6 of the Workshop #SMM4H (Social Media Mining for Health) 2025, specifically focusing on potentially identifying personally experienced vaccine adverse reactions from unstructured social media posts. We present an enhanced hybrid methodology that combines the sophisticated chain-of-thought (CoT) reasoning of GPT-4o and the enhancement of domain-specific knowledge with pre-trained language models (PLM) such as BERTweet-large and DeBERTa-v3-base. By systematically employing stratified cross-validation, robust regularization (mixout and layer-wise learning rate decay), and conditional text enhancement, our model significantly outperforms the previous benchmarks, achieving an F1 score of 0.96 on both validation and test sets, underscoring the effectiveness of integrating human-like reasoning capabilities with traditional classification models.

PDF