From Visual Metaphor to Social Resonance: Evaluating Multimodal Large Language Models in Internet Meme Communication

Internet memes, as quintessential units of digital culture, rely heavily on visual metaphors to convey complex social sentiments. This study investigates the capacity of Multimodal Large Language Models (MLLMs) to decode these implicit symbols and predicts their potential to trigger social resonance. Drawing upon Lakoff's Conceptual Metaphor Theory and Collins' Interaction Ritual chains provide a cross-disciplinary framework to evaluate the "computational empathy" of models like GPT-4o and Claude 3.5. We conducted a two-stage experiment: first, a decoding benchmark using Chain-of-Thought (CoT) prompting across 500 stratified meme samples; second, a statistical fitting analysis correlating AI-generated sentiment scores with real-world engagement metrics from social media platforms. Key findings indicate that: the introduction of CoT significantly improves irony detection from 45% to 82%, suggesting that MLLM resonance is a product of logical simulation rather than biological intuition. A significant "Cultural Alignment Bias" exists, where Western-centric training data leads to semantic collapse when decoding localized metaphors. Statistical fitting reveals a non-linear "U-shaped" relationship between semantic clarity and social resonance; memes within a "semantic ambiguity zone" elicit the deepest user interaction. This research delineates the boundaries of algorithmic sentiment alignment, provides a novel "semantic-logic" pathway for digital governance and public opinion forecasting.