E-GuidedRE: An Evaluation-Model-Guided Collaborative Framework for Relation Extraction in Specialized Domains

Relation Extraction is crucial for knowledge graph construction, but extracting complex relations in specialized domains like Satellite Navigation Countermeasures (SNCM) remains challenging due to long semantic spans and high relational density. While Large Language Models (LLMs) possess strong semantic understanding, they often suffer from severe recall deficiency and hallucinations in high-density multi-entity contexts. Conversely, traditional small models generate excessive redundant noise. To address these limitations, this paper proposes an evaluation-model-guided relation extraction method (E-guidedRE). This framework employs a two-stage collaborative mechanism. First, a lightweight evaluation model utilizing a GlobalPointer network with Rotary Position Embedding (RoPE) and a sparse multi-label loss function acts as a structural filter to generate high-coverage candidate entity pairs. Second, these candidates guide the frozen LLM to perform deep semantic discrimination and retrospective denoising. Furthermore, we construct a dedicated SNCM dataset to fill the vertical domain data void. Extensive experiments across five heterogeneous datasets, including general, biomedical, financial, and our self-built SNCM corpus, demonstrate that E-guidedRE exhibits remarkable robustness. In ablation studies on the SNCM dataset, our method improved the F1-score from 36.54% to 54.93% compared to standalone LLM extraction, boosting recall from 27.81% to 47.13%. The proposed paradigm effectively mitigates the LLM’s attention divergence in complex contexts, dynamically balancing precision and recall, and offers a highly reliable technical pathway for knowledge extraction in specialized vertical domains.