Engineering Implementation and Experimental Research of Video Structured Tag Generation System Based on Multimodal Information Fusion

With the rapid development of short video platforms and User Generated Content (UGC), video data has accounted for an increasing proportion in internet information dissemination. Automatically extracting computable, retrievable, and interpretable structured semantic tags from massive unstructured videos has become a key technical issue in content recommendation, information retrieval, and content moderation systems. This paper proposes and implements an engineering-reproducible multimodal structured tag generation system for videos. Adopting a modular multimodal fusion architecture, the system extracts semantic information from three modalities—visual, speech, and on-screen text. Through unified structured representation and multimodal Prompt design, multi-source information is input into a large language model for joint reasoning, ultimately generating JSON-format output including topic hierarchy, multi-tag attributes, and confidence estimation. Experimental results show that the multimodal fusion method significantly outperforms single-modal schemes in topic classification and attribute recognition tasks. Moreover, the system can stably output structured JSON results, demonstrating high engineering usability. This paper details the engineering implementation experience and key technical trade-offs of the system under the Windows environment, providing a reference for subsequent research and practical deployment.