Structure-aware parsing method for coal mine standard documents

To enhance the efficiency of knowledge supply and the accuracy of compliance determination for coal mine safety relevant standard documents in intelligent applications, and to address the issues of semantic representation drift and lack of logical boundaries caused by physical chunking in traditional Retrieval-Augmented Generation (RAG) systems, this study proposes and constructs a Structure-Aware Logical Chunking (SALC) algorithm. First, based on a fine-tuned Layout Language Model version 3 (LayoutLMv3) model, the method integrates text, layout, and image features to perform fine-grained semantic annotation on documents. Second, a semantic conflict self-correcting document topology parsing algorithm is designed to construct a Document Tree with Path Semantics (DTPS). Finally, a semantic retrieval tag-driven dynamic aggregation strategy is employed to generate structurally complete logical chunks. Experimental results show that the model achieves a Macro F₁ score of 93.53% and a Clause-Level F₁ score of 91.79%, outperforming Bidirectional Encoder Representations from Transformers (BERT) and Document Understanding Transformer (Donut). In end-to-end retrieval evaluation, the SALC algorithm achieves a Hit Rate of 93.36%, a Mean Reciprocal Rank (MRR) of 0.89, and a Context Precision of 71.92%, surpassing the Recursive Character Text Splitter (RCTS) and the Query Decomposition driven iterative Retrieval Augmented Generation (QD-RAG) model. Compared with the coal mine domain document parsing models of Coal Mine Safety Equipment Knowledge Graph Construction Method (CMSE-KGM) and Prompt Relation oriented Parsing Engine-Universal Information Extraction (PRoPE-UIE), SALC improves the Hit Rate by 3.60 and 6.94 percentage points respectively, the MRR by 0.09 and 0.06. Results regarding different document lengths indicate that SALC maintains a Hit Rate of 82.17% for long documents exceeding 20 pages and scenarios with Optical Character Recognition (OCR) noise interference, significantly outperforming comparative models. Ablation studies further validate the effectiveness of the topology self-correction and dynamic aggregation strategies.