Abstract: Named entity recognition (NER) of wheat pests and diseases is essential for extracting structured information from agricultural texts and constructing knowledge graphs, thereby supporting intelligent pest management and reducing economic losses in wheat production. With the rapid accumulation of unstructured agricultural texts, field investigation reports, and pest management guidelines, accurate and efficient NER has become increasingly important. However, two major challenges remain. First, publicly available, high-quality annotated corpora for wheat pests and diseases are scarce, while manual annotation requires substantial domain expertise and labor, limiting the availability of training data. Second, conventional NER methods often struggle to identify nested entities and determine entity boundaries in complex agricultural texts, which can compromise recognition accuracy and generalization. To address these challenges, this study constructed a high-quality annotated corpus of wheat pests and diseases and developed a collaborative annotation framework integrating large language models (LLMs) with manual calibration. By combining the semantic understanding capabilities of LLMs with human verification, the framework aims to improve annotation quality and alleviate data scarcity in low-resource agricultural settings. Furthermore, an NER model, termed MB-BSG (MacBERT-BiGRU-multi-head self-attention-global pointer), was developed. Specifically, MacBERT was used as the pre-trained language model to obtain contextualized semantic representations of Chinese agricultural texts. A bidirectional gated recurrent unit (BiGRU) was integrated with a multi-head self-attention mechanism to capture sequential dependencies and local contextual features while modeling long-range semantic interactions. A Global Pointer network was then employed to identify entity spans, enabling the recognition of nested entities and more precise entity boundary localization. Experimental results showed that the LLM-assisted collaborative annotation framework achieved an average F1 score of 89.52%, exceeding manual annotation and weak supervision by 1.20 and 1.77 percentage points, respectively, while effectively reducing annotation noise. On the constructed corpus, MB-BSG achieved precision, recall, and F1 scores of 91.57%, 90.08%, and 90.82%, respectively, outperforming the conventional and baseline models evaluated in this study. Efficiency analysis further showed that MB-BSG contained 108.17 million parameters and required 22.80 GFLOPs, with a single-sentence inference time of 69.01 ms and peak memory consumption of 451.83 MB. These results indicate that the model achieves a favorable balance between recognition performance and computational cost, suggesting its potential for deployment in resource-constrained environments. Overall, this study provides a high-quality annotated corpus and an efficient NER approach for wheat pest and disease knowledge extraction, offering methodological support for knowledge graph construction and precision pest management in intelligent agriculture.