A UIE-based entity recognition and brief generation agent for energy policy texts

In response to energy policy texts under the "dual-carbon" strategy, which exhibit high-frequency production, rapidupdates, and dense terminology, this study proposes an "extraction-generation" closed-loop framework to addresslimitations of traditional methods in handling long entities and cross-sentence consistency. On the upstreamside, wepropose UIE-PolicyAlign based on the Unified Information Extraction framework, employing dual-pointer spanmodeling and template alignment to effectively handle boundary drift and cross-sentence fragmentation, achieving 2.9% improvement in precision over UIE-Base. On the downstream side, we design a three-layer brief-generation Agent that adopts controlled generation strategy, filling confirmed structured information first, then invoking large language modelsunder constraints while maintaining traceability through source anchors. The system incorporates a quality-drivendegradation mechanism for robustness. Experiments on 6,292 energy policy documents demonstrate significant reduction in factual hallucinations compared to pure generation methods, with ROUGE-1 F score of 70.22% and 55% higher processing efficiency. This study constructs a transferable and auditable paradigm for information extractionandbrief generation, providing reliable data foundation for policy compliance assessment and decision-making support.