Contrastive Learning-Based Semantic Recognition Model for Cross-Departmental Documents in Public Organizations

作者：H H Li, Zhilin Guo, Zijun Zhang · 年份：2026 · DOI：10.1109/cnml68938.2026.11452416 · 研究领域：Advanced Graph Neural Networks、E-Government and Public Services、Text and Document Classification Technologies

Cross-departmental collaborative governance in modern digital government faces bottlenecks including terminology islands, administrative caliber drift, and implicit logical conflicts. To address these issues, we propose MGCF-Gov, a multi-granularity contrastive fusion network that learns an isomorphic mapping from unstructured public appeals to structured government rules. The framework integrates (i) a prompt-driven encoder that injects administrative-domain priors into a pre-trained backbone, (ii) a Multi-granularity Contrastive Learning (MGCL) objective that aligns representations at both macro-intent and micro-terminology levels, and (iii) a Dynamic Semantic Fusion (DSF) module that performs cross-source evidence aggregation via attention-based fusion. Since contrastive objectives can suffer from representation collapse, we introduce explicit collapse diagnostics (e.g., embedding variance, effective rank, and alignment/uniformity metrics) and analyze how MGCL mitigates collapse risk. Experiments on MGCF-Gov-Set (750,000 real-world samples) show that MGCF-Gov achieves an F1-score of 0.832 and an AUC of 0.944, outperforming strong neural baselines such as RoBERTa-wwm and RGCN under the same evaluation protocol.