A Prompt-Tuned Large Language Model Framework for Automated Generation of Structured Information

Large Language Models (LLMs) have achieved remarkable success in natural language understanding and generation. However, generating structured information from unstructured text remains challenging due to format instability, low extraction accuracy, poor generalization, and the high cost of model fine-tuning. In practical applications, these issues cause difficulties in accurately extracting structured information, especially limiting performance in cross-domain and complex scenarios. To address these issues, we propose PT4Struct (Prompt Tuning for Structured Information), a unified framework that leverages prompt tuning to guide LLMs in automatically generating structured outputs such as tables, JSON objects, and key-value pairs. Instead of modifying model parameters, PT4Struct designs task-specific prompts to steer the model toward format-compliant outputs, significantly reducing training and deployment costs. The framework exhibits strong generalizability and extensibility across various structured information generation tasks. Experimental results demonstrate that PT4Struct outperforms existing methods in accuracy, controllability, and cross-task adaptability, offering an efficient, flexible, and cost-effective solution for structured data generation. This method provides an efficient and generalizable solution for multi-domain structured information extraction, with potential for extension to more complex text types in future work.