Abstract:In order to alleviate label noise and error propagation caused by the entity co-occurrence assumption in distantly supervised annotation of agricultural pest and disease texts, an evidence-constrained distant supervision-based annotation denoising method is proposed. In this method, entity co-occurrence results obtained by distant supervision are used as candidate triple recall signals. Agricultural knowledge graph Schema, field-structured textual evidence and graph evidence are then integrated to verify candidate triples under evidence constraints, and a large language model is employed to generate initial annotations with evidence support. On this basis, a class-prototype-guided representation enhancement mechanism is introduced, and an uncertainty evaluation strategy is used to selectively write back labels to be added or removed, thereby realizing iterative refinement of noisy annotation sets. The experimental results show that, under the same CasRel extractor, the F1 score of relation triple extraction is improved from 0.631 to 0.832 by evidence-constrained initial annotation, and further reaches 0.897 when combined with the self-learning denoising mechanism. Compared with original distant supervision annotation, the proposed method effectively improves annotation quality and relation extraction performance. It is concluded that this method can reduce annotation noise in agricultural pest and disease texts and provide high-quality data support for agricultural pest and disease knowledge graph construction.