Mark-KG: A Feedback-Driven Prompting Framework for Marketing Knowledge Graphs

Abstract

Due to rapid growth of unstructured digital information generated across multiple platforms, marketing intelligence has evolved into a highly data-intense domain. Although Large Language Models (LLMs) have proved strong performance when it comes to zero-shot information extraction tasks, their practical deployment in complex enterprise issues possess a serious problem due to issues such as hallucination and the absence of persistent contextual memory. In the presented work, a novel architecture is presented which enhances marketing intelligence through the integration of an autonomous Knowledge Graph (KG) with Adaptive Prompt Engineering (APE). The proposed system on its own determines the best extraction strategy by taking into consideration the historical data of performance patterns along with the available graph context. As the complexity of the input depends, the framework can adapt in between different prompting approaches, ranging from zero-shot techniques to graph augmented prompting methods. As the KG expands, it continuously provides the ground for future LLM prompts, which in short improves the quality over time. This enhancement mechanism contributes towards faster sample efficiency as well as greater extraction accuracy. The system was tested across different benchmarks, and the results indicate that the proposed APE+KG framework achieves a mean F1-score of 0.625 which outperforms rule-based and Chain-of-Thought (CoT) baselines by reducing the computational latency by half. The results obtained establish a scalable and efficiency foundation for the development of AI-driven enterprise intelligence systems.