HSTM: Decoding Humor in Memes Through Multimodal Intelligence

Humor plays an essential role in human communication, impacting emotions, social interaction, and cultural articulation. In the digital age, memes have arisen as a prevailing form of humorous expression by imaginatively combining text and images. Nevertheless, understanding humor in memes is difficult owing to their inherently multimodal structure and the frequent use of sarcasm, irony, and cultural context. Standard sentiment analysis methods often rely on single-modal inputs and find it difficult to capture this complexity. To tackle these limitations, we suggest the humor-specific multimodal transformer (HSMT) a deep learning architecture that merges visual and textual features using an attention-based fusion mechanism. The model incorporates a bidirectional encoder representation from transformers (BERT)-based text encoder and a ResNet-Based image encoder to capture subtle linguistic and visual hints. Furthermore, we incorporate a dedicated sarcasm detection module to boost interpretation of sentiment conveyed in reverse or ironic ways. Evaluated on the Memotion dataset, HSMT shows significant enhancements in classifying humor-related sentiment compared to baseline models. The model demonstrates practical applicability in areas such as social media monitoring, content moderation, brand analysis, and intelligent conversational systems. Future directions involve multilingual humor detection, real-time meme analysis, and adaptive learning strategies to deal with the changing nature of humor.