Structured Extraction and BPMN Generation From Chinese Procedural Texts via Hybrid Attention-Based Domain Language Models

Understanding and modeling workflows is crucial for advancing digital transformation and intelligent process management in the electric power industry. While recent progress in large language models (LLMs) has significantly enhanced natural language processing capabilities, data in the power sector presents unique challenges—including complex formatting, limited structured annotations, and extensive use of domain-specific terminology—that hinder the direct application of general-purpose LLMs. Moreover, conventional attention mechanisms often struggle with long procedural texts due to high computational overhead, excessive resource consumption, and unstable global semantic representations, further limiting their effectiveness in power workflow modeling tasks. To address these challenges, this paper proposes a domain-adapted LLM framework tailored for the standardization of electric power workflows and introduces a specialized model, Lumina BPM, designed for this task. The model incorporates a hybrid sparse self-attention mechanism that retains attention on initial tokens and dynamically selects the most relevant historical tokens based on query context. This design reduces the need to compute full-sequence dense attention matrices, thereby lowering computational costs while improving the model’s ability to capture critical information from lengthy and complex process descriptions. Experimental results show that Lumina BPM accurately transforms standardized procedural texts into executable piperflow statements and generates corresponding BPMN 2.0 diagrams. The proposed approach significantly improves both the accuracy and interpretability of workflow modeling, offering a practical pathway toward intelligent process standardization in critical infrastructure domains.