The spread of misinformation across online social networks poses a significant threat to political stability, societal harmony, and economic integrity. This paper tackles the challenge of analyzing online posts, which often consist of diverse heterogeneous modalities like text or image. We introduce a multimodal architecture designed to accurately identify manipulated content. Our approach leverages a large-scale, multimodal dataset encompassing images, captions, comments, and metadata for each post. A novel encoding strategy is employed to capture both the semantic content and hierarchical structure of the comments. Based on CLIP, an early fusion technique is applied to process and merge the hidden representations. This method effectively combines multimodal and unimodal data according to their respective information channels. The performance is evaluated against the Fakeddit dataset, achieving a competent accuracy in binary classification tasks (0.9506 acctest) with significant hardware limitations, and outperforming the SotA in more complex categorization tasks, with 0.9509 acctest and 0.9371 acctest in 3-way and 6-way label classification, respectively.