Optimizing multimodal achievement retrieval algorithms for large-scale technology markets

To enhance retrieval quality for multimodal scientific achievements in large-scale technology markets, this study constructs a unified representation framework for heterogeneous text, images, and structured metrics. It introduces crossmodal semantic alignment, dynamic weight learning, modality importance discrimination, and heterogeneous data augmentation strategies to improve the model's adaptability to modality differences and imbalance. This study investigates alignment methods for multimodal features within a high-dimensional unified semantic space and validates the contribution of different optimization modules to retrieval performance through comparative experiments. Results demonstrate that the proposed multimodal retrieval method exhibits significant advantages in both recall and relevance metrics, effectively supporting semantic matching requirements for complex achievement information.