A Transformer-Based Multimodal Semantic Retrieval Model for Business Intelligence Systems

Abstract

In the increasingly growing business intelligence (BI) environment of multi-source heterogeneous data, traditional information retrieval methods face significant bottlenecks in accuracy, response efficiency, and semantic understanding ability. We aim to investigate whether multimodal semantic modeling and dynamic intent recognition can significantly improve retrieval precision and response efficiency in BI contexts. This paper designs and implements a Transformer-based multimodal semantic retrieval model architecture, which combines a multi-layer semantic modeling mechanism with a context enhancement strategy to model the deep matching relationship between user queries and multimodal business data. The architecture introduces a query semantic vector generation module based on Transformer encoders, adopts a multi-channel deep feature fusion structure for structured fields, behavior logs, and documents, and incorporates a dynamic user intent recognition module for context-aware representation. The training employs a contrastive loss with softmax normalization, optimized with the AdamW optimizer and cosine learning rate scheduling. Experiments are conducted on three enterprise-level datasets, including an internal document corpus (42,000+ samples), a structured product dataset (18,000 records), and user behavior logs (3.1M entries). Evaluation results demonstrate that the proposed model outperforms BM25, DSSM, and BERT Retriever, achieving Precision@10 = 0.723, nDCG@10 = 0.702, and MRR = 0.537, with relative improvements of up to 28.6%. In addition, the model reduces average response latency to 430 ms and maintains a user satisfaction score above 87, proving its feasibility for deployment in intelligent decision-support BI platforms.