A Real-Time Meta-Discourse Dynamic Graph Construction and Purchase Intention Decoding Model Based on a Multimodal Transformer Architecture

Meta-Discourse plays a critical role in shaping consumer decisions, and its multimodal mechanisms and real-time optimization needs have become key topics at the intersection of AI and consumer behavior research. To address the limitations of unimodal analysis, opaque decision-making processes, and delayed feedback, this paper proposes a real-time Meta-Discourse dynamic graph construction and purchase intention decoding model based on a multimodal Transformer architecture. The model integrates text, audio, and visual features via a hierarchical Transformer encoder for cross-modal semantic alignment. A temporal graph neural network captures the dynamic evolution of metadiscursive strategies, while a neuro-symbolic fusion decoder translates purchase intention into interpretable dimensions such as trust, emotional arousal, and brand awareness. With lightweight deployment, the model supports real-time strategy optimization within 200 milliseconds. Experiments on the TikTok livestreaming dataset show that the model achieves an 85.7% accuracy in purchase intentions recognition, which is 23.6% higher than baseline models. Additionally, a novel metric, the Meta-Discourse Cognitive Conduction Index (MCC-I), is proposed for strategy evaluation. This paper offers a practical, real-time multimodal decision support system for livestream marketing and advances AI-driven consumer behavior analysis.