Real or perceived corruption can have a damaging effect on health care services and outcomes. In particular, research suggests perceived corruption had a significant impact on COVID-19 vaccination. Given the role of social media in health communications, identifying and understanding perceived corruption related to vaccines and vaccination is critical to build societal cohesion and public trust in health institutions and strategies, manage and combat misinformation and disinformation, and design more effective policies, interventions, and communications strategies. There is a dearth of research on binary and multi-class classification of corruption dialogues in health or otherwise. We address this gap by introducing a general hierarchical corruption dialogue taxonomy (HCDT) and formulating binary and multi-class classification tasks based on the HCDT. We also create a vaccine-specific labelled dataset for each task, and fine-tune three large language models (BERT, RoBERTa, and BERTweet) based on these datasets. We evaluate the performance of these models in the binary and multi-class classification tasks. While all models performed similarly for the binary task, RoBERTa performed best for multi-class classification of corruption dialogue.