Boosting Movie and TV Tag Accuracy with Knowledge Graphs

Abstract

This paper introduces a fine-grained video tag classification algorithm that integrates knowledge graphs to identify specific TV series. A multimodal pretraining model extracts visual and textual features, feeding them into a multitask prediction model to classify type, genre, and entity tags. A similarity task enhances feature compactness, while an entity correction model refines predictions using co-occurrence data from knowledge graphs. Using a knowledge graph built from the data of Douban, a leading platform in China, the model achieved a 3.70% improvement in Top-1 accuracy for type tags, 3.35% for genre tags, and 16.57% for entity tags. The global-local attention model further increased entity tag accuracy from 38.7% to 45.6%. This approach boosts classification accuracy and provides insights for handling limited data.