Construction and Application of a Multimodal Knowledge Graph for Social Media Opinion Analysis Based on Large Language Models

The rapid expansion of multimodal content on social media brings both prospects and challenges to public opinion analysis. While such data provides deep insights into collective sentiment and societal dynamics, it also heightens the risks of misinformation spread and cognitive overload for analytical models. To tackle these issues, this paper presents a new approach to constructing a multimodal knowledge graph enabled by large language models, aiming to boost reasoning, traceability, and verifiability in social media opinion analysis. The proposed architecture incorporates LLM-enhanced multimodal semantic alignment, event-centric temporal relation extraction, and cross-modal confidence calibration to construct dynamic knowledge graphs for evolving public discourse. Experiments on a curated dataset of real-world events, which includes adversarial samples with inserted logical inconsistencies, demonstrate that our method outperforms traditional RAG approaches. Specifically, it achieves a 23.9 % improvement in F1 score for information tracing and an 8.0 % increase in accuracy for misinformation detection. These results validate the framework's capacity to reduce information overload and identify latent logical inconsistencies. However, challenges remain in real-time adaptation and resolving deep semantic contradictions, indicating future paths in combining dynamic network structures and domain-specific knowledge for early misinformation containment.